* filestore flusher = false , correct my problem of constant write (need info on this parameter) [not found] <48b5f93e-fe36-458f-a05e-9f9fd2293249@mailpro> @ 2012-06-20 5:20 ` Alexandre DERUMIER 2012-06-21 16:03 ` Sage Weil 0 siblings, 1 reply; 11+ messages in thread From: Alexandre DERUMIER @ 2012-06-20 5:20 UTC (permalink / raw) To: ceph-devel; +Cc: Mark Nelson, Stefan Priebe Hi, I have tried to disabe filestore flusher filestore flusher = false filestore max sync interval = 30 filestore min sync interval = 29 in osd config. now, I see correct sync each 30s when doing rados bench rados -p pool3 bench 60 write -t 16 seekwatcher movie: before ------ http://odisoweb1.odiso.net/seqwrite-radosbench-flusherenable.mpg after ----- http://odisoweb1.odiso.net/seqwrite-radosbench-flusherdisable.mpg Shouldn't it be the normal behaviour ? What's exactly is filestore flusher vs syncfs ? This seem to works fine with rados bench, But when I launch benchmark with fio from my guest vm, I see again constant write. (I'll try to debug that today) My target is to be able to handle small random write and write them each 30s. Regards, Alexandre ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-20 5:20 ` filestore flusher = false , correct my problem of constant write (need info on this parameter) Alexandre DERUMIER @ 2012-06-21 16:03 ` Sage Weil 2012-06-22 6:14 ` Alexandre DERUMIER 0 siblings, 1 reply; 11+ messages in thread From: Sage Weil @ 2012-06-21 16:03 UTC (permalink / raw) To: Alexandre DERUMIER; +Cc: ceph-devel, Mark Nelson, Stefan Priebe Hi Alexandre, [Sorry I didn't follow up earlier; I didn't understand your question.] If you turn off the journal compeletely, you will see bursty write commits from the perspective of the client, because the OSD is periodically doing a sync or snapshot and only acking the writes then. If you enable the journal, the OSD will reply with a commit as soon as the write is stable in the journal. That's one reason why it is there--file system commits of heavyweight and slow. If we left the file system to its own devices and did a sync every 10 seconds, the disk would sit idle while a bunch of dirty data accumulated in cache, and then the sync/snapshot would take a really long time. This is horribly inefficient (the disk is idle half the time), and useless (the delayed write behavior makes sense for local workloads, but not servers where there is a client on the other end batching its writes). To prevent this, 'filestore flusher' will prod the kernel to flush out any written data to the disk quickly. Then, when we get around to doing the sync/snapshot it is pretty quick, because only fs metadata and just-written data needs to be flushed. So: the behavior you're seeing is normal, and good. Did I understand your confusion correctly? Thanks! sage On Wed, 20 Jun 2012, Alexandre DERUMIER wrote: > Hi, > I have tried to disabe filestore flusher > > filestore flusher = false > filestore max sync interval = 30 > filestore min sync interval = 29 > > > in osd config. > > > now, I see correct sync each 30s when doing rados bench > > rados -p pool3 bench 60 write -t 16 > > > seekwatcher movie: > > > before > ------ > http://odisoweb1.odiso.net/seqwrite-radosbench-flusherenable.mpg > > after > ----- > http://odisoweb1.odiso.net/seqwrite-radosbench-flusherdisable.mpg > > > Shouldn't it be the normal behaviour ? What's exactly is filestore flusher vs syncfs ? > > > > This seem to works fine with rados bench, > But when I launch benchmark with fio from my guest vm, I see again constant write. > (I'll try to debug that today) > > > My target is to be able to handle small random write and write them each 30s. > > Regards, > > Alexandre > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-21 16:03 ` Sage Weil @ 2012-06-22 6:14 ` Alexandre DERUMIER 2012-06-23 15:38 ` Sage Weil 0 siblings, 1 reply; 11+ messages in thread From: Alexandre DERUMIER @ 2012-06-22 6:14 UTC (permalink / raw) To: Sage Weil; +Cc: ceph-devel, Mark Nelson, Stefan Priebe Hi Sage, thanks for your response. >>If you turn off the journal compeletely, you will see bursty write commits >>from the perspective of the client, because the OSD is periodically doing >>a sync or snapshot and only acking the writes then. >>If you enable the journal, the OSD will reply with a commit as soon as the >>write is stable in the journal. That's one reason why it is there--file >>system commits of heavyweight and slow. Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. >>If we left the file system to its own devices and did a sync every 10 >>seconds, the disk would sit idle while a bunch of dirty data accumulated >>in cache, and then the sync/snapshot would take a really long time. This >>is horribly inefficient (the disk is idle half the time), and useless (the >>delayed write behavior makes sense for local workloads, but not servers >>where there is a client on the other end batching its writes). To prevent >>this, 'filestore flusher' will prod the kernel to flush out any written >>data to the disk quickly. Then, when we get around to doing the >>sync/snapshot it is pretty quick, because only fs metadata and >>just-written data needs to be flushed. mmm, I disagree. If you flush quickly, it's works fine with sequential write workload. But if you have a lot of random write with 4k block by exemple, you are going to have a lot of disk seeks. The way zfs works or netapp san storage works, they take random writes in a fast journal then flush them sequentially each 20s to slow storage. To compare with zfs or netapp, I can achieve around 20000io/s on random write 4K with 4GB nvram and 10 x 7200 disk. with ceph, i'm around 2000io/s with same config. (3 nodes with 10x7200disk, 2x replication), so around real disk io limit without any write cache. So for now, i'm think i'm going to use ssd for my osds,I have 80% random write workload. (no seeks, so no problem to constant random write) NTW: maybe wiki is wrong http://ceph.com/wiki/OSD_journal section Motivation "Enterprise products like NetApp filers "cheat" by journaling all writes to NVRAM and then taking their time to flush things out to disk efficiently. This gives you very low-latency writes _and_ efficient disk IO at the expense of hardware." This why I thinked ceph worked like this. Thanks again, -Alexandre ----- Mail original ----- De: "Sage Weil" <sage@inktank.com> À: "Alexandre DERUMIER" <aderumier@odiso.com> Cc: ceph-devel@vger.kernel.org, "Mark Nelson" <mark.nelson@inktank.com>, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Jeudi 21 Juin 2012 18:03:45 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) Hi Alexandre, [Sorry I didn't follow up earlier; I didn't understand your question.] If you turn off the journal compeletely, you will see bursty write commits from the perspective of the client, because the OSD is periodically doing a sync or snapshot and only acking the writes then. If you enable the journal, the OSD will reply with a commit as soon as the write is stable in the journal. That's one reason why it is there--file system commits of heavyweight and slow. If we left the file system to its own devices and did a sync every 10 seconds, the disk would sit idle while a bunch of dirty data accumulated in cache, and then the sync/snapshot would take a really long time. This is horribly inefficient (the disk is idle half the time), and useless (the delayed write behavior makes sense for local workloads, but not servers where there is a client on the other end batching its writes). To prevent this, 'filestore flusher' will prod the kernel to flush out any written data to the disk quickly. Then, when we get around to doing the sync/snapshot it is pretty quick, because only fs metadata and just-written data needs to be flushed. So: the behavior you're seeing is normal, and good. Did I understand your confusion correctly? Thanks! sage On Wed, 20 Jun 2012, Alexandre DERUMIER wrote: > Hi, > I have tried to disabe filestore flusher > > filestore flusher = false > filestore max sync interval = 30 > filestore min sync interval = 29 > > > in osd config. > > > now, I see correct sync each 30s when doing rados bench > > rados -p pool3 bench 60 write -t 16 > > > seekwatcher movie: > > > before > ------ > http://odisoweb1.odiso.net/seqwrite-radosbench-flusherenable.mpg > > after > ----- > http://odisoweb1.odiso.net/seqwrite-radosbench-flusherdisable.mpg > > > Shouldn't it be the normal behaviour ? What's exactly is filestore flusher vs syncfs ? > > > > This seem to works fine with rados bench, > But when I launch benchmark with fio from my guest vm, I see again constant write. > (I'll try to debug that today) > > > My target is to be able to handle small random write and write them each 30s. > > Regards, > > Alexandre > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-22 6:14 ` Alexandre DERUMIER @ 2012-06-23 15:38 ` Sage Weil 2012-06-23 16:40 ` Mark Nelson 0 siblings, 1 reply; 11+ messages in thread From: Sage Weil @ 2012-06-23 15:38 UTC (permalink / raw) To: Alexandre DERUMIER; +Cc: ceph-devel, Mark Nelson, Stefan Priebe [-- Attachment #1: Type: TEXT/PLAIN, Size: 6900 bytes --] On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: > Hi Sage, > thanks for your response. > > >>If you turn off the journal compeletely, you will see bursty write commits > >>from the perspective of the client, because the OSD is periodically doing > >>a sync or snapshot and only acking the writes then. > >>If you enable the journal, the OSD will reply with a commit as soon as the > >>write is stable in the journal. That's one reason why it is there--file > >>system commits of heavyweight and slow. > > Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. > > >>If we left the file system to its own devices and did a sync every 10 > >>seconds, the disk would sit idle while a bunch of dirty data accumulated > >>in cache, and then the sync/snapshot would take a really long time. This > >>is horribly inefficient (the disk is idle half the time), and useless (the > >>delayed write behavior makes sense for local workloads, but not servers > >>where there is a client on the other end batching its writes). To prevent > >>this, 'filestore flusher' will prod the kernel to flush out any written > >>data to the disk quickly. Then, when we get around to doing the > >>sync/snapshot it is pretty quick, because only fs metadata and > >>just-written data needs to be flushed. > > mmm, I disagree. > > If you flush quickly, it's works fine with sequential write workload. > > But if you have a lot of random write with 4k block by exemple, you are > going to have a lot of disk seeks. The way zfs works or netapp san > storage works, they take random writes in a fast journal then flush them > sequentially each 20s to slow storage. Oh, I see what you're getting at. Yes, that is not ideal for small random writes. There is a branch in ceph.git called wip-flushmin that just sets a minimum write size for the flush that will probably do a decent job of dealing with this: small writes won't get flushed, large ones will. Picking the right value will depend on how expensive seeks are for your storage system. You'll want to cherry-pick just the top commit on top of whatever it is you're running... sage > > To compare with zfs or netapp, I can achieve around 20000io/s on random > write 4K with 4GB nvram and 10 x 7200 disk. > > with ceph, i'm around 2000io/s with same config. (3 nodes with > 10x7200disk, 2x replication), so around real disk io limit without any > write cache. > > > So for now, i'm think i'm going to use ssd for my osds,I have 80% random > write workload. (no seeks, so no problem to constant random write) > > > > NTW: maybe wiki is wrong > http://ceph.com/wiki/OSD_journal > section Motivation > "Enterprise products like NetApp filers "cheat" by journaling all writes to NVRAM and then taking their time to flush things out to disk efficiently. This gives you very low-latency writes _and_ efficient disk IO at the expense of hardware." > > This why I thinked ceph worked like this. > > > Thanks again, > > -Alexandre > > > > > > > > ----- Mail original ----- > > De: "Sage Weil" <sage@inktank.com> > À: "Alexandre DERUMIER" <aderumier@odiso.com> > Cc: ceph-devel@vger.kernel.org, "Mark Nelson" <mark.nelson@inktank.com>, "Stefan Priebe" <s.priebe@profihost.ag> > Envoyé: Jeudi 21 Juin 2012 18:03:45 > Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) > > Hi Alexandre, > > [Sorry I didn't follow up earlier; I didn't understand your question.] > > If you turn off the journal compeletely, you will see bursty write commits > from the perspective of the client, because the OSD is periodically doing > a sync or snapshot and only acking the writes then. > > If you enable the journal, the OSD will reply with a commit as soon as the > write is stable in the journal. That's one reason why it is there--file > system commits of heavyweight and slow. > > If we left the file system to its own devices and did a sync every 10 > seconds, the disk would sit idle while a bunch of dirty data accumulated > in cache, and then the sync/snapshot would take a really long time. This > is horribly inefficient (the disk is idle half the time), and useless (the > delayed write behavior makes sense for local workloads, but not servers > where there is a client on the other end batching its writes). To prevent > this, 'filestore flusher' will prod the kernel to flush out any written > data to the disk quickly. Then, when we get around to doing the > sync/snapshot it is pretty quick, because only fs metadata and > just-written data needs to be flushed. > > So: the behavior you're seeing is normal, and good. > > Did I understand your confusion correctly? > > Thanks! > sage > > > On Wed, 20 Jun 2012, Alexandre DERUMIER wrote: > > > Hi, > > I have tried to disabe filestore flusher > > > > filestore flusher = false > > filestore max sync interval = 30 > > filestore min sync interval = 29 > > > > > > in osd config. > > > > > > now, I see correct sync each 30s when doing rados bench > > > > rados -p pool3 bench 60 write -t 16 > > > > > > seekwatcher movie: > > > > > > before > > ------ > > http://odisoweb1.odiso.net/seqwrite-radosbench-flusherenable.mpg > > > > after > > ----- > > http://odisoweb1.odiso.net/seqwrite-radosbench-flusherdisable.mpg > > > > > > Shouldn't it be the normal behaviour ? What's exactly is filestore flusher vs syncfs ? > > > > > > > > This seem to works fine with rados bench, > > But when I launch benchmark with fio from my guest vm, I see again constant write. > > (I'll try to debug that today) > > > > > > My target is to be able to handle small random write and write them each 30s. > > > > Regards, > > > > Alexandre > > -- > > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > > the body of a message to majordomo@vger.kernel.org > > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > -- > > -- > > > > > > Alexandre D e rumier > > Ingénieur Systèmes et Réseaux > > > Fixe : 03 20 68 88 85 > > Fax : 03 20 68 90 88 > > > 45 Bvd du Général Leclerc 59100 Roubaix > 12 rue Marivaux 75002 Paris > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-23 15:38 ` Sage Weil @ 2012-06-23 16:40 ` Mark Nelson 2012-06-23 17:46 ` Alexandre DERUMIER 0 siblings, 1 reply; 11+ messages in thread From: Mark Nelson @ 2012-06-23 16:40 UTC (permalink / raw) To: Sage Weil; +Cc: Alexandre DERUMIER, ceph-devel, Stefan Priebe On 6/23/12 10:38 AM, Sage Weil wrote: > On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: >> Hi Sage, >> thanks for your response. >> >>>> If you turn off the journal compeletely, you will see bursty write commits >>> >from the perspective of the client, because the OSD is periodically doing >>>> a sync or snapshot and only acking the writes then. >>>> If you enable the journal, the OSD will reply with a commit as soon as the >>>> write is stable in the journal. That's one reason why it is there--file >>>> system commits of heavyweight and slow. >> >> Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. >> >>>> If we left the file system to its own devices and did a sync every 10 >>>> seconds, the disk would sit idle while a bunch of dirty data accumulated >>>> in cache, and then the sync/snapshot would take a really long time. This >>>> is horribly inefficient (the disk is idle half the time), and useless (the >>>> delayed write behavior makes sense for local workloads, but not servers >>>> where there is a client on the other end batching its writes). To prevent >>>> this, 'filestore flusher' will prod the kernel to flush out any written >>>> data to the disk quickly. Then, when we get around to doing the >>>> sync/snapshot it is pretty quick, because only fs metadata and >>>> just-written data needs to be flushed. >> >> mmm, I disagree. >> >> If you flush quickly, it's works fine with sequential write workload. >> >> But if you have a lot of random write with 4k block by exemple, you are >> going to have a lot of disk seeks. The way zfs works or netapp san >> storage works, they take random writes in a fast journal then flush them >> sequentially each 20s to slow storage. > > Oh, I see what you're getting at. Yes, that is not ideal for small random > writes. There is a branch in ceph.git called wip-flushmin that just sets > a minimum write size for the flush that will probably do a decent job of > dealing with this: small writes won't get flushed, large ones will. > Picking the right value will depend on how expensive seeks are for your > storage system. > > You'll want to cherry-pick just the top commit on top of whatever it is > you're running... I was just talking with Elder on IRC yesterday about looking into how much small network transfers are hurting us in cases like these. Even with SSD based OSDs I haven't seen a very dramatic improvement in small request performance. How tough would it be to aggregate requests into larger network transactions? There would be a latency penalty of course, but we could flush a client side dirty cache pretty quickly and still benefit if we are getting bombarded with lots of tiny requests. Mark ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-23 16:40 ` Mark Nelson @ 2012-06-23 17:46 ` Alexandre DERUMIER 2012-06-23 18:12 ` Sage Weil 0 siblings, 1 reply; 11+ messages in thread From: Alexandre DERUMIER @ 2012-06-23 17:46 UTC (permalink / raw) To: Mark Nelson; +Cc: ceph-devel, Stefan Priebe, Sage Weil >>I was just talking with Elder on IRC yesterday about looking into how >>much small network transfers are hurting us in cases like these. Even >>with SSD based OSDs I haven't seen a very dramatic improvement in small >>request performance. How tough would it be to aggregate requests into >>larger network transactions? There would be a latency penalty of >>course, but we could flush a client side dirty cache pretty quickly and >>still benefit if we are getting bombarded with lots of tiny requests. Yes, I see no improvement with journal on tmpfs ...this is strange.. Also, I have tried with rbd_cache=true, so ios should be already aggregate in bigger transaction. But I didnt't have see any improvement. I'm around 2000 ios. Do you know what is the bottleneck ? rbd protocol (some kind of overhead for each io ?) ----- Mail original ----- De: "Mark Nelson" <mark.nelson@inktank.com> À: "Sage Weil" <sage@inktank.com> Cc: "Alexandre DERUMIER" <aderumier@odiso.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 18:40:28 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) On 6/23/12 10:38 AM, Sage Weil wrote: > On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: >> Hi Sage, >> thanks for your response. >> >>>> If you turn off the journal compeletely, you will see bursty write commits >>> >from the perspective of the client, because the OSD is periodically doing >>>> a sync or snapshot and only acking the writes then. >>>> If you enable the journal, the OSD will reply with a commit as soon as the >>>> write is stable in the journal. That's one reason why it is there--file >>>> system commits of heavyweight and slow. >> >> Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. >> >>>> If we left the file system to its own devices and did a sync every 10 >>>> seconds, the disk would sit idle while a bunch of dirty data accumulated >>>> in cache, and then the sync/snapshot would take a really long time. This >>>> is horribly inefficient (the disk is idle half the time), and useless (the >>>> delayed write behavior makes sense for local workloads, but not servers >>>> where there is a client on the other end batching its writes). To prevent >>>> this, 'filestore flusher' will prod the kernel to flush out any written >>>> data to the disk quickly. Then, when we get around to doing the >>>> sync/snapshot it is pretty quick, because only fs metadata and >>>> just-written data needs to be flushed. >> >> mmm, I disagree. >> >> If you flush quickly, it's works fine with sequential write workload. >> >> But if you have a lot of random write with 4k block by exemple, you are >> going to have a lot of disk seeks. The way zfs works or netapp san >> storage works, they take random writes in a fast journal then flush them >> sequentially each 20s to slow storage. > > Oh, I see what you're getting at. Yes, that is not ideal for small random > writes. There is a branch in ceph.git called wip-flushmin that just sets > a minimum write size for the flush that will probably do a decent job of > dealing with this: small writes won't get flushed, large ones will. > Picking the right value will depend on how expensive seeks are for your > storage system. > > You'll want to cherry-pick just the top commit on top of whatever it is > you're running... I was just talking with Elder on IRC yesterday about looking into how much small network transfers are hurting us in cases like these. Even with SSD based OSDs I haven't seen a very dramatic improvement in small request performance. How tough would it be to aggregate requests into larger network transactions? There would be a latency penalty of course, but we could flush a client side dirty cache pretty quickly and still benefit if we are getting bombarded with lots of tiny requests. Mark -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-23 17:46 ` Alexandre DERUMIER @ 2012-06-23 18:12 ` Sage Weil 2012-06-23 18:21 ` Alexandre DERUMIER 0 siblings, 1 reply; 11+ messages in thread From: Sage Weil @ 2012-06-23 18:12 UTC (permalink / raw) To: Alexandre DERUMIER; +Cc: Mark Nelson, ceph-devel, Stefan Priebe [-- Attachment #1: Type: TEXT/PLAIN, Size: 5300 bytes --] On Sat, 23 Jun 2012, Alexandre DERUMIER wrote: > >>I was just talking with Elder on IRC yesterday about looking into how > >>much small network transfers are hurting us in cases like these. Even > >>with SSD based OSDs I haven't seen a very dramatic improvement in small > >>request performance. How tough would it be to aggregate requests into > >>larger network transactions? There would be a latency penalty of > >>course, but we could flush a client side dirty cache pretty quickly and > >>still benefit if we are getting bombarded with lots of tiny requests. > > Yes, I see no improvement with journal on tmpfs ...this is strange.. > > Also, I have tried with rbd_cache=true, so ios should be already aggregate in bigger transaction. > But I didnt't have see any improvement. > > I'm around 2000 ios. > > Do you know what is the bottleneck ? rbd protocol (some kind of overhead > for each io ?) Is that 2000 ios from a single client? You might try multiple clients and see if the sum of the ios will scale any higher. That will tell us whether it is in the messenger or osd request pipeline. The latter definitely needs some work, although there may be a quick fix to the msgr that will buy us some too. sage > > > ----- Mail original ----- > > De: "Mark Nelson" <mark.nelson@inktank.com> > À: "Sage Weil" <sage@inktank.com> > Cc: "Alexandre DERUMIER" <aderumier@odiso.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> > Envoyé: Samedi 23 Juin 2012 18:40:28 > Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) > > On 6/23/12 10:38 AM, Sage Weil wrote: > > On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: > >> Hi Sage, > >> thanks for your response. > >> > >>>> If you turn off the journal compeletely, you will see bursty write commits > >>> >from the perspective of the client, because the OSD is periodically doing > >>>> a sync or snapshot and only acking the writes then. > >>>> If you enable the journal, the OSD will reply with a commit as soon as the > >>>> write is stable in the journal. That's one reason why it is there--file > >>>> system commits of heavyweight and slow. > >> > >> Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. > >> > >>>> If we left the file system to its own devices and did a sync every 10 > >>>> seconds, the disk would sit idle while a bunch of dirty data accumulated > >>>> in cache, and then the sync/snapshot would take a really long time. This > >>>> is horribly inefficient (the disk is idle half the time), and useless (the > >>>> delayed write behavior makes sense for local workloads, but not servers > >>>> where there is a client on the other end batching its writes). To prevent > >>>> this, 'filestore flusher' will prod the kernel to flush out any written > >>>> data to the disk quickly. Then, when we get around to doing the > >>>> sync/snapshot it is pretty quick, because only fs metadata and > >>>> just-written data needs to be flushed. > >> > >> mmm, I disagree. > >> > >> If you flush quickly, it's works fine with sequential write workload. > >> > >> But if you have a lot of random write with 4k block by exemple, you are > >> going to have a lot of disk seeks. The way zfs works or netapp san > >> storage works, they take random writes in a fast journal then flush them > >> sequentially each 20s to slow storage. > > > > Oh, I see what you're getting at. Yes, that is not ideal for small random > > writes. There is a branch in ceph.git called wip-flushmin that just sets > > a minimum write size for the flush that will probably do a decent job of > > dealing with this: small writes won't get flushed, large ones will. > > Picking the right value will depend on how expensive seeks are for your > > storage system. > > > > You'll want to cherry-pick just the top commit on top of whatever it is > > you're running... > > I was just talking with Elder on IRC yesterday about looking into how > much small network transfers are hurting us in cases like these. Even > with SSD based OSDs I haven't seen a very dramatic improvement in small > request performance. How tough would it be to aggregate requests into > larger network transactions? There would be a latency penalty of > course, but we could flush a client side dirty cache pretty quickly and > still benefit if we are getting bombarded with lots of tiny requests. > > Mark > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > -- > > -- > > > > > > Alexandre D e rumier > > Ingénieur Systèmes et Réseaux > > > Fixe : 03 20 68 88 85 > > Fax : 03 20 68 90 88 > > > 45 Bvd du Général Leclerc 59100 Roubaix > 12 rue Marivaux 75002 Paris > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-23 18:12 ` Sage Weil @ 2012-06-23 18:21 ` Alexandre DERUMIER 2012-06-24 8:10 ` Alexandre DERUMIER 0 siblings, 1 reply; 11+ messages in thread From: Alexandre DERUMIER @ 2012-06-23 18:21 UTC (permalink / raw) To: Sage Weil; +Cc: Mark Nelson, ceph-devel, Stefan Priebe >>Is that 2000 ios from a single client? You might try multiple clients and >>see if the sum of the ios will scale any higher. yes from a single client. (qemu-kvm guest). Tomorrow,I'll retry with 3 qemu-kvm guest, on same host and 3 differents hosts. I'll also try on bigger cpu machine to compra.(I see a lot of cpu on my kvm guest process, more than with iscsi) I'll keep you in touch. Thanks Alexandre ----- Mail original ----- De: "Sage Weil" <sage@inktank.com> À: "Alexandre DERUMIER" <aderumier@odiso.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 20:12:49 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) On Sat, 23 Jun 2012, Alexandre DERUMIER wrote: > >>I was just talking with Elder on IRC yesterday about looking into how > >>much small network transfers are hurting us in cases like these. Even > >>with SSD based OSDs I haven't seen a very dramatic improvement in small > >>request performance. How tough would it be to aggregate requests into > >>larger network transactions? There would be a latency penalty of > >>course, but we could flush a client side dirty cache pretty quickly and > >>still benefit if we are getting bombarded with lots of tiny requests. > > Yes, I see no improvement with journal on tmpfs ...this is strange.. > > Also, I have tried with rbd_cache=true, so ios should be already aggregate in bigger transaction. > But I didnt't have see any improvement. > > I'm around 2000 ios. > > Do you know what is the bottleneck ? rbd protocol (some kind of overhead > for each io ?) Is that 2000 ios from a single client? You might try multiple clients and see if the sum of the ios will scale any higher. That will tell us whether it is in the messenger or osd request pipeline. The latter definitely needs some work, although there may be a quick fix to the msgr that will buy us some too. sage > > > ----- Mail original ----- > > De: "Mark Nelson" <mark.nelson@inktank.com> > À: "Sage Weil" <sage@inktank.com> > Cc: "Alexandre DERUMIER" <aderumier@odiso.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> > Envoyé: Samedi 23 Juin 2012 18:40:28 > Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) > > On 6/23/12 10:38 AM, Sage Weil wrote: > > On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: > >> Hi Sage, > >> thanks for your response. > >> > >>>> If you turn off the journal compeletely, you will see bursty write commits > >>> >from the perspective of the client, because the OSD is periodically doing > >>>> a sync or snapshot and only acking the writes then. > >>>> If you enable the journal, the OSD will reply with a commit as soon as the > >>>> write is stable in the journal. That's one reason why it is there--file > >>>> system commits of heavyweight and slow. > >> > >> Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. > >> > >>>> If we left the file system to its own devices and did a sync every 10 > >>>> seconds, the disk would sit idle while a bunch of dirty data accumulated > >>>> in cache, and then the sync/snapshot would take a really long time. This > >>>> is horribly inefficient (the disk is idle half the time), and useless (the > >>>> delayed write behavior makes sense for local workloads, but not servers > >>>> where there is a client on the other end batching its writes). To prevent > >>>> this, 'filestore flusher' will prod the kernel to flush out any written > >>>> data to the disk quickly. Then, when we get around to doing the > >>>> sync/snapshot it is pretty quick, because only fs metadata and > >>>> just-written data needs to be flushed. > >> > >> mmm, I disagree. > >> > >> If you flush quickly, it's works fine with sequential write workload. > >> > >> But if you have a lot of random write with 4k block by exemple, you are > >> going to have a lot of disk seeks. The way zfs works or netapp san > >> storage works, they take random writes in a fast journal then flush them > >> sequentially each 20s to slow storage. > > > > Oh, I see what you're getting at. Yes, that is not ideal for small random > > writes. There is a branch in ceph.git called wip-flushmin that just sets > > a minimum write size for the flush that will probably do a decent job of > > dealing with this: small writes won't get flushed, large ones will. > > Picking the right value will depend on how expensive seeks are for your > > storage system. > > > > You'll want to cherry-pick just the top commit on top of whatever it is > > you're running... > > I was just talking with Elder on IRC yesterday about looking into how > much small network transfers are hurting us in cases like these. Even > with SSD based OSDs I haven't seen a very dramatic improvement in small > request performance. How tough would it be to aggregate requests into > larger network transactions? There would be a latency penalty of > course, but we could flush a client side dirty cache pretty quickly and > still benefit if we are getting bombarded with lots of tiny requests. > > Mark > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > -- > > -- > > > > > > Alexandre D e rumier > > Ingénieur Systèmes et Réseaux > > > Fixe : 03 20 68 88 85 > > Fax : 03 20 68 90 88 > > > 45 Bvd du Général Leclerc 59100 Roubaix > 12 rue Marivaux 75002 Paris > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-23 18:21 ` Alexandre DERUMIER @ 2012-06-24 8:10 ` Alexandre DERUMIER 2012-06-25 14:34 ` Alexandre DERUMIER 0 siblings, 1 reply; 11+ messages in thread From: Alexandre DERUMIER @ 2012-06-24 8:10 UTC (permalink / raw) To: Sage Weil; +Cc: Mark Nelson, ceph-devel, Stefan Priebe ok, I have done tests with more than 1 client. 3 kvm guest on 3 differents kvm host server and 3 kvm guest on same server. I have the same result , around 2000 io/s shared between clients. So it doesn't scale. I have also tried with 3 nodes x 5 osd 15k drive + 5 tmpfs journal and 3 nodes x 1 osd (hardware raid0 with 5 x 15kdisk) + 1 tmpfs journal the results are same, around 2000 io/s But if I try with 3 nodes x 1 osd with 1x15k drive, I got around 500 io/s I also known that stefan priebe has achieve around 12000io/s with ssds in osd. So it's seem related to osd drive speed So, are we sure that journal is acking to client before flushing to disk ? benchmark used is fio, with directio, write 100MB of 4K block. (so journal is big enough to handle all the write) fio --filename=[disk] --direct=1 --rw=randwrite --bs=4k --size=100M --numjobs=50 --runtime=30 --group_reporting --name=file1 ----- Mail original ----- De: "Alexandre DERUMIER" <aderumier@odiso.com> À: "Sage Weil" <sage@inktank.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 20:21:05 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) >>Is that 2000 ios from a single client? You might try multiple clients and >>see if the sum of the ios will scale any higher. yes from a single client. (qemu-kvm guest). Tomorrow,I'll retry with 3 qemu-kvm guest, on same host and 3 differents hosts. I'll also try on bigger cpu machine to compra.(I see a lot of cpu on my kvm guest process, more than with iscsi) I'll keep you in touch. Thanks Alexandre ----- Mail original ----- De: "Sage Weil" <sage@inktank.com> À: "Alexandre DERUMIER" <aderumier@odiso.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 20:12:49 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) On Sat, 23 Jun 2012, Alexandre DERUMIER wrote: > >>I was just talking with Elder on IRC yesterday about looking into how > >>much small network transfers are hurting us in cases like these. Even > >>with SSD based OSDs I haven't seen a very dramatic improvement in small > >>request performance. How tough would it be to aggregate requests into > >>larger network transactions? There would be a latency penalty of > >>course, but we could flush a client side dirty cache pretty quickly and > >>still benefit if we are getting bombarded with lots of tiny requests. > > Yes, I see no improvement with journal on tmpfs ...this is strange.. > > Also, I have tried with rbd_cache=true, so ios should be already aggregate in bigger transaction. > But I didnt't have see any improvement. > > I'm around 2000 ios. > > Do you know what is the bottleneck ? rbd protocol (some kind of overhead > for each io ?) Is that 2000 ios from a single client? You might try multiple clients and see if the sum of the ios will scale any higher. That will tell us whether it is in the messenger or osd request pipeline. The latter definitely needs some work, although there may be a quick fix to the msgr that will buy us some too. sage > > > ----- Mail original ----- > > De: "Mark Nelson" <mark.nelson@inktank.com> > À: "Sage Weil" <sage@inktank.com> > Cc: "Alexandre DERUMIER" <aderumier@odiso.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> > Envoyé: Samedi 23 Juin 2012 18:40:28 > Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) > > On 6/23/12 10:38 AM, Sage Weil wrote: > > On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: > >> Hi Sage, > >> thanks for your response. > >> > >>>> If you turn off the journal compeletely, you will see bursty write commits > >>> >from the perspective of the client, because the OSD is periodically doing > >>>> a sync or snapshot and only acking the writes then. > >>>> If you enable the journal, the OSD will reply with a commit as soon as the > >>>> write is stable in the journal. That's one reason why it is there--file > >>>> system commits of heavyweight and slow. > >> > >> Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. > >> > >>>> If we left the file system to its own devices and did a sync every 10 > >>>> seconds, the disk would sit idle while a bunch of dirty data accumulated > >>>> in cache, and then the sync/snapshot would take a really long time. This > >>>> is horribly inefficient (the disk is idle half the time), and useless (the > >>>> delayed write behavior makes sense for local workloads, but not servers > >>>> where there is a client on the other end batching its writes). To prevent > >>>> this, 'filestore flusher' will prod the kernel to flush out any written > >>>> data to the disk quickly. Then, when we get around to doing the > >>>> sync/snapshot it is pretty quick, because only fs metadata and > >>>> just-written data needs to be flushed. > >> > >> mmm, I disagree. > >> > >> If you flush quickly, it's works fine with sequential write workload. > >> > >> But if you have a lot of random write with 4k block by exemple, you are > >> going to have a lot of disk seeks. The way zfs works or netapp san > >> storage works, they take random writes in a fast journal then flush them > >> sequentially each 20s to slow storage. > > > > Oh, I see what you're getting at. Yes, that is not ideal for small random > > writes. There is a branch in ceph.git called wip-flushmin that just sets > > a minimum write size for the flush that will probably do a decent job of > > dealing with this: small writes won't get flushed, large ones will. > > Picking the right value will depend on how expensive seeks are for your > > storage system. > > > > You'll want to cherry-pick just the top commit on top of whatever it is > > you're running... > > I was just talking with Elder on IRC yesterday about looking into how > much small network transfers are hurting us in cases like these. Even > with SSD based OSDs I haven't seen a very dramatic improvement in small > request performance. How tough would it be to aggregate requests into > larger network transactions? There would be a latency penalty of > course, but we could flush a client side dirty cache pretty quickly and > still benefit if we are getting bombarded with lots of tiny requests. > > Mark > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > -- > > -- > > > > > > Alexandre D e rumier > > Ingénieur Systèmes et Réseaux > > > Fixe : 03 20 68 88 85 > > Fax : 03 20 68 90 88 > > > 45 Bvd du Général Leclerc 59100 Roubaix > 12 rue Marivaux 75002 Paris > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-24 8:10 ` Alexandre DERUMIER @ 2012-06-25 14:34 ` Alexandre DERUMIER 2012-06-26 7:29 ` Alexandre DERUMIER 0 siblings, 1 reply; 11+ messages in thread From: Alexandre DERUMIER @ 2012-06-25 14:34 UTC (permalink / raw) To: Sage Weil; +Cc: Mark Nelson, ceph-devel, Stefan Priebe Hi, maybe this can help: I have tunned filestore queue max ops = 50000 now I'm able to achieve 4000io/s (with some spikes) with 3 nodes with 1 x osd (1 x 15k drive by osd), journal on tmpfs or 3 nodes with 5 osd (1 x 15k drive by osd), journal on tmpfs same result for both conf. ----- Mail original ----- De: "Alexandre DERUMIER" <aderumier@odiso.com> À: "Sage Weil" <sage@inktank.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Dimanche 24 Juin 2012 10:10:48 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) ok, I have done tests with more than 1 client. 3 kvm guest on 3 differents kvm host server and 3 kvm guest on same server. I have the same result , around 2000 io/s shared between clients. So it doesn't scale. I have also tried with 3 nodes x 5 osd 15k drive + 5 tmpfs journal and 3 nodes x 1 osd (hardware raid0 with 5 x 15kdisk) + 1 tmpfs journal the results are same, around 2000 io/s But if I try with 3 nodes x 1 osd with 1x15k drive, I got around 500 io/s I also known that stefan priebe has achieve around 12000io/s with ssds in osd. So it's seem related to osd drive speed So, are we sure that journal is acking to client before flushing to disk ? benchmark used is fio, with directio, write 100MB of 4K block. (so journal is big enough to handle all the write) fio --filename=[disk] --direct=1 --rw=randwrite --bs=4k --size=100M --numjobs=50 --runtime=30 --group_reporting --name=file1 ----- Mail original ----- De: "Alexandre DERUMIER" <aderumier@odiso.com> À: "Sage Weil" <sage@inktank.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 20:21:05 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) >>Is that 2000 ios from a single client? You might try multiple clients and >>see if the sum of the ios will scale any higher. yes from a single client. (qemu-kvm guest). Tomorrow,I'll retry with 3 qemu-kvm guest, on same host and 3 differents hosts. I'll also try on bigger cpu machine to compra.(I see a lot of cpu on my kvm guest process, more than with iscsi) I'll keep you in touch. Thanks Alexandre ----- Mail original ----- De: "Sage Weil" <sage@inktank.com> À: "Alexandre DERUMIER" <aderumier@odiso.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 20:12:49 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) On Sat, 23 Jun 2012, Alexandre DERUMIER wrote: > >>I was just talking with Elder on IRC yesterday about looking into how > >>much small network transfers are hurting us in cases like these. Even > >>with SSD based OSDs I haven't seen a very dramatic improvement in small > >>request performance. How tough would it be to aggregate requests into > >>larger network transactions? There would be a latency penalty of > >>course, but we could flush a client side dirty cache pretty quickly and > >>still benefit if we are getting bombarded with lots of tiny requests. > > Yes, I see no improvement with journal on tmpfs ...this is strange.. > > Also, I have tried with rbd_cache=true, so ios should be already aggregate in bigger transaction. > But I didnt't have see any improvement. > > I'm around 2000 ios. > > Do you know what is the bottleneck ? rbd protocol (some kind of overhead > for each io ?) Is that 2000 ios from a single client? You might try multiple clients and see if the sum of the ios will scale any higher. That will tell us whether it is in the messenger or osd request pipeline. The latter definitely needs some work, although there may be a quick fix to the msgr that will buy us some too. sage > > > ----- Mail original ----- > > De: "Mark Nelson" <mark.nelson@inktank.com> > À: "Sage Weil" <sage@inktank.com> > Cc: "Alexandre DERUMIER" <aderumier@odiso.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> > Envoyé: Samedi 23 Juin 2012 18:40:28 > Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) > > On 6/23/12 10:38 AM, Sage Weil wrote: > > On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: > >> Hi Sage, > >> thanks for your response. > >> > >>>> If you turn off the journal compeletely, you will see bursty write commits > >>> >from the perspective of the client, because the OSD is periodically doing > >>>> a sync or snapshot and only acking the writes then. > >>>> If you enable the journal, the OSD will reply with a commit as soon as the > >>>> write is stable in the journal. That's one reason why it is there--file > >>>> system commits of heavyweight and slow. > >> > >> Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. > >> > >>>> If we left the file system to its own devices and did a sync every 10 > >>>> seconds, the disk would sit idle while a bunch of dirty data accumulated > >>>> in cache, and then the sync/snapshot would take a really long time. This > >>>> is horribly inefficient (the disk is idle half the time), and useless (the > >>>> delayed write behavior makes sense for local workloads, but not servers > >>>> where there is a client on the other end batching its writes). To prevent > >>>> this, 'filestore flusher' will prod the kernel to flush out any written > >>>> data to the disk quickly. Then, when we get around to doing the > >>>> sync/snapshot it is pretty quick, because only fs metadata and > >>>> just-written data needs to be flushed. > >> > >> mmm, I disagree. > >> > >> If you flush quickly, it's works fine with sequential write workload. > >> > >> But if you have a lot of random write with 4k block by exemple, you are > >> going to have a lot of disk seeks. The way zfs works or netapp san > >> storage works, they take random writes in a fast journal then flush them > >> sequentially each 20s to slow storage. > > > > Oh, I see what you're getting at. Yes, that is not ideal for small random > > writes. There is a branch in ceph.git called wip-flushmin that just sets > > a minimum write size for the flush that will probably do a decent job of > > dealing with this: small writes won't get flushed, large ones will. > > Picking the right value will depend on how expensive seeks are for your > > storage system. > > > > You'll want to cherry-pick just the top commit on top of whatever it is > > you're running... > > I was just talking with Elder on IRC yesterday about looking into how > much small network transfers are hurting us in cases like these. Even > with SSD based OSDs I haven't seen a very dramatic improvement in small > request performance. How tough would it be to aggregate requests into > larger network transactions? There would be a latency penalty of > course, but we could flush a client side dirty cache pretty quickly and > still benefit if we are getting bombarded with lots of tiny requests. > > Mark > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > -- > > -- > > > > > > Alexandre D e rumier > > Ingénieur Systèmes et Réseaux > > > Fixe : 03 20 68 88 85 > > Fax : 03 20 68 90 88 > > > 45 Bvd du Général Leclerc 59100 Roubaix > 12 rue Marivaux 75002 Paris > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html ^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) 2012-06-25 14:34 ` Alexandre DERUMIER @ 2012-06-26 7:29 ` Alexandre DERUMIER 0 siblings, 0 replies; 11+ messages in thread From: Alexandre DERUMIER @ 2012-06-26 7:29 UTC (permalink / raw) To: Sage Weil; +Cc: Mark Nelson, ceph-devel, Stefan Priebe some tests today: with filestore queue max ops = 50000 I can achieve 4000io/s BY kvm-guest. So it seem to scale. (tested with 3 kvm guest, each guest 4000io/s). I have some spike because underlying storage is not able to handle write. with iostat and seekwatcher I see a lot of random write and seeks on osd storage. (I'll take alook to flush_min wip to see if it's helping). ----- Mail original ----- De: "Alexandre DERUMIER" <aderumier@odiso.com> À: "Sage Weil" <sage@inktank.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Lundi 25 Juin 2012 16:34:03 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) Hi, maybe this can help: I have tunned filestore queue max ops = 50000 now I'm able to achieve 4000io/s (with some spikes) with 3 nodes with 1 x osd (1 x 15k drive by osd), journal on tmpfs or 3 nodes with 5 osd (1 x 15k drive by osd), journal on tmpfs same result for both conf. ----- Mail original ----- De: "Alexandre DERUMIER" <aderumier@odiso.com> À: "Sage Weil" <sage@inktank.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Dimanche 24 Juin 2012 10:10:48 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) ok, I have done tests with more than 1 client. 3 kvm guest on 3 differents kvm host server and 3 kvm guest on same server. I have the same result , around 2000 io/s shared between clients. So it doesn't scale. I have also tried with 3 nodes x 5 osd 15k drive + 5 tmpfs journal and 3 nodes x 1 osd (hardware raid0 with 5 x 15kdisk) + 1 tmpfs journal the results are same, around 2000 io/s But if I try with 3 nodes x 1 osd with 1x15k drive, I got around 500 io/s I also known that stefan priebe has achieve around 12000io/s with ssds in osd. So it's seem related to osd drive speed So, are we sure that journal is acking to client before flushing to disk ? benchmark used is fio, with directio, write 100MB of 4K block. (so journal is big enough to handle all the write) fio --filename=[disk] --direct=1 --rw=randwrite --bs=4k --size=100M --numjobs=50 --runtime=30 --group_reporting --name=file1 ----- Mail original ----- De: "Alexandre DERUMIER" <aderumier@odiso.com> À: "Sage Weil" <sage@inktank.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 20:21:05 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) >>Is that 2000 ios from a single client? You might try multiple clients and >>see if the sum of the ios will scale any higher. yes from a single client. (qemu-kvm guest). Tomorrow,I'll retry with 3 qemu-kvm guest, on same host and 3 differents hosts. I'll also try on bigger cpu machine to compra.(I see a lot of cpu on my kvm guest process, more than with iscsi) I'll keep you in touch. Thanks Alexandre ----- Mail original ----- De: "Sage Weil" <sage@inktank.com> À: "Alexandre DERUMIER" <aderumier@odiso.com> Cc: "Mark Nelson" <mark.nelson@inktank.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> Envoyé: Samedi 23 Juin 2012 20:12:49 Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) On Sat, 23 Jun 2012, Alexandre DERUMIER wrote: > >>I was just talking with Elder on IRC yesterday about looking into how > >>much small network transfers are hurting us in cases like these. Even > >>with SSD based OSDs I haven't seen a very dramatic improvement in small > >>request performance. How tough would it be to aggregate requests into > >>larger network transactions? There would be a latency penalty of > >>course, but we could flush a client side dirty cache pretty quickly and > >>still benefit if we are getting bombarded with lots of tiny requests. > > Yes, I see no improvement with journal on tmpfs ...this is strange.. > > Also, I have tried with rbd_cache=true, so ios should be already aggregate in bigger transaction. > But I didnt't have see any improvement. > > I'm around 2000 ios. > > Do you know what is the bottleneck ? rbd protocol (some kind of overhead > for each io ?) Is that 2000 ios from a single client? You might try multiple clients and see if the sum of the ios will scale any higher. That will tell us whether it is in the messenger or osd request pipeline. The latter definitely needs some work, although there may be a quick fix to the msgr that will buy us some too. sage > > > ----- Mail original ----- > > De: "Mark Nelson" <mark.nelson@inktank.com> > À: "Sage Weil" <sage@inktank.com> > Cc: "Alexandre DERUMIER" <aderumier@odiso.com>, ceph-devel@vger.kernel.org, "Stefan Priebe" <s.priebe@profihost.ag> > Envoyé: Samedi 23 Juin 2012 18:40:28 > Objet: Re: filestore flusher = false , correct my problem of constant write (need info on this parameter) > > On 6/23/12 10:38 AM, Sage Weil wrote: > > On Fri, 22 Jun 2012, Alexandre DERUMIER wrote: > >> Hi Sage, > >> thanks for your response. > >> > >>>> If you turn off the journal compeletely, you will see bursty write commits > >>> >from the perspective of the client, because the OSD is periodically doing > >>>> a sync or snapshot and only acking the writes then. > >>>> If you enable the journal, the OSD will reply with a commit as soon as the > >>>> write is stable in the journal. That's one reason why it is there--file > >>>> system commits of heavyweight and slow. > >> > >> Yes of course, I don't wan't to desactivate journal, using a journal on a fast ssd or nvram is the right way. > >> > >>>> If we left the file system to its own devices and did a sync every 10 > >>>> seconds, the disk would sit idle while a bunch of dirty data accumulated > >>>> in cache, and then the sync/snapshot would take a really long time. This > >>>> is horribly inefficient (the disk is idle half the time), and useless (the > >>>> delayed write behavior makes sense for local workloads, but not servers > >>>> where there is a client on the other end batching its writes). To prevent > >>>> this, 'filestore flusher' will prod the kernel to flush out any written > >>>> data to the disk quickly. Then, when we get around to doing the > >>>> sync/snapshot it is pretty quick, because only fs metadata and > >>>> just-written data needs to be flushed. > >> > >> mmm, I disagree. > >> > >> If you flush quickly, it's works fine with sequential write workload. > >> > >> But if you have a lot of random write with 4k block by exemple, you are > >> going to have a lot of disk seeks. The way zfs works or netapp san > >> storage works, they take random writes in a fast journal then flush them > >> sequentially each 20s to slow storage. > > > > Oh, I see what you're getting at. Yes, that is not ideal for small random > > writes. There is a branch in ceph.git called wip-flushmin that just sets > > a minimum write size for the flush that will probably do a decent job of > > dealing with this: small writes won't get flushed, large ones will. > > Picking the right value will depend on how expensive seeks are for your > > storage system. > > > > You'll want to cherry-pick just the top commit on top of whatever it is > > you're running... > > I was just talking with Elder on IRC yesterday about looking into how > much small network transfers are hurting us in cases like these. Even > with SSD based OSDs I haven't seen a very dramatic improvement in small > request performance. How tough would it be to aggregate requests into > larger network transactions? There would be a latency penalty of > course, but we could flush a client side dirty cache pretty quickly and > still benefit if we are getting bombarded with lots of tiny requests. > > Mark > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > -- > > -- > > > > > > Alexandre D e rumier > > Ingénieur Systèmes et Réseaux > > > Fixe : 03 20 68 88 85 > > Fax : 03 20 68 90 88 > > > 45 Bvd du Général Leclerc 59100 Roubaix > 12 rue Marivaux 75002 Paris > -- > To unsubscribe from this list: send the line "unsubscribe ceph-devel" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html > > -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html -- -- Alexandre D e rumier Ingénieur Systèmes et Réseaux Fixe : 03 20 68 88 85 Fax : 03 20 68 90 88 45 Bvd du Général Leclerc 59100 Roubaix 12 rue Marivaux 75002 Paris -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html ^ permalink raw reply [flat|nested] 11+ messages in thread
end of thread, other threads:[~2012-06-26 7:29 UTC | newest]
Thread overview: 11+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
[not found] <48b5f93e-fe36-458f-a05e-9f9fd2293249@mailpro>
2012-06-20 5:20 ` filestore flusher = false , correct my problem of constant write (need info on this parameter) Alexandre DERUMIER
2012-06-21 16:03 ` Sage Weil
2012-06-22 6:14 ` Alexandre DERUMIER
2012-06-23 15:38 ` Sage Weil
2012-06-23 16:40 ` Mark Nelson
2012-06-23 17:46 ` Alexandre DERUMIER
2012-06-23 18:12 ` Sage Weil
2012-06-23 18:21 ` Alexandre DERUMIER
2012-06-24 8:10 ` Alexandre DERUMIER
2012-06-25 14:34 ` Alexandre DERUMIER
2012-06-26 7:29 ` Alexandre DERUMIER
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.