[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Jie Liu]

Hi All,

These days, I am investigating an issue regarding OCFS2 unexpected
reboot in some real world use cases.
This problem occurred when the network status goes south, when the disk
IO load is too high, etc...
I suspect it might caused by ocfs2 fencing if it's BIO reading/writing
can not be scheduled and processed quickly, or
something like this happened in the network IO heartbeat thread.

Now am trying to reproduce this problem locally.  In the meantime, I'd
like to ping you guys with some rough ideas
to improve the disk IO heartbeat to see if they are sounds reasonable or
not.

Firstly, if an OCFS2 node is suffer from heavy disk IO, how about to fix
the bio read/write to make this IO request can not
be preempted by other requests? e.g,  for o2hb_issue_node_write(),
currently, it do bio submission with WRITE only,
'submit_bio(WRITE, bio)'.   If we change the flag to WRITE_SYNC, or even
submit the request combine with REQ_FUA,
maybe could get highest priority for disk IO request.

Secondly, the comments for bio allocation at o2hb_setup_one_bio()
indicates that we can pre-allocate bio instead of
acquire for each time.  But I have not saw any code snippet doing such
things in kernel. :(
how about creating a private bio set for each o2hb_region, so that we
can do allocation out of it?
maybe it's faster than do allocation from global bio sets.  Also, does
it make sense if creating a memory pool
on each o2hb_region, so that we can have continuous pages bind to those
bios?


Any comments are appreciated!

Thanks,
-Jeff






-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120822/c454d0b4/attachment.html 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[srinivas eeda]

On 8/22/2012 7:17 AM, Jie Liu wrote:
> Hi All,
>
> These days, I am investigating an issue regarding OCFS2 unexpected 
> reboot in some real world use cases.
> This problem occurred when the network status goes south, when the 
> disk IO load is too high, etc...
> I suspect it might caused by ocfs2 fencing if it's BIO reading/writing 
> can not be scheduled and processed quickly, or
> something like this happened in the network IO heartbeat thread.
>
> Now am trying to reproduce this problem locally.  In the meantime, I'd 
> like to ping you guys with some rough ideas
> to improve the disk IO heartbeat to see if they are sounds reasonable 
> or not.
>
> Firstly, if an OCFS2 node is suffer from heavy disk IO, how about to 
> fix the bio read/write to make this IO request can not
> be preempted by other requests? e.g,  for o2hb_issue_node_write(), 
> currently, it do bio submission with WRITE only,
> 'submit_bio(WRITE, bio)'.   If we change the flag to WRITE_SYNC, or 
> even submit the request combine with REQ_FUA,
> maybe could get highest priority for disk IO request.
This was submitted before by Noboru Iwamatsu and acked by sunil and tao 
but some how didn't get merged

https://oss.oracle.com/pipermail/ocfs2-devel/2011-December/008438.html
>
> Secondly, the comments for bio allocation at o2hb_setup_one_bio() 
> indicates that we can pre-allocate bio instead of
> acquire for each time.  But I have not saw any code snippet doing such 
> things in kernel. :(
> how about creating a private bio set for each o2hb_region, so that we 
> can do allocation out of it?
> maybe it's faster than do allocation from global bio sets.  Also, does 
> it make sense if creating a memory pool
> on each o2hb_region, so that we can have continuous pages bind to 
> those bios?
>
>
> Any comments are appreciated!
>
> Thanks,
> -Jeff
>
>
>
>
>
>
>
>
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120822/7967b991/attachment-0001.html 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Sunil Mushran]

Yes. WRITE_SYNC should be good. Not FUA.

Also, you may want to look into using io priorities. The code is all there.
Just needs activation.

On Wed, Aug 22, 2012 at 10:13 AM, srinivas eeda <srinivas.eeda@oracle.com>wrote:

>
> On 8/22/2012 7:17 AM, Jie Liu wrote:
>
> Hi All,
>
> These days, I am investigating an issue regarding OCFS2 unexpected reboot
> in some real world use cases.
> This problem occurred when the network status goes south, when the disk IO
> load is too high, etc...
> I suspect it might caused by ocfs2 fencing if it's BIO reading/writing can
> not be scheduled and processed quickly, or
> something like this happened in the network IO heartbeat thread.
>
> Now am trying to reproduce this problem locally.  In the meantime, I'd
> like to ping you guys with some rough ideas
> to improve the disk IO heartbeat to see if they are sounds reasonable or
> not.
>
> Firstly, if an OCFS2 node is suffer from heavy disk IO, how about to fix
> the bio read/write to make this IO request can not
> be preempted by other requests? e.g,  for o2hb_issue_node_write(),
> currently, it do bio submission with WRITE only,
> 'submit_bio(WRITE, bio)'.   If we change the flag to WRITE_SYNC, or even
> submit the request combine with REQ_FUA,
> maybe could get highest priority for disk IO request.
>
> This was submitted before by Noboru Iwamatsu and acked by sunil and tao
> but some how didn't get merged
>
> https://oss.oracle.com/pipermail/ocfs2-devel/2011-December/008438.html
>
>
> Secondly, the comments for bio allocation at o2hb_setup_one_bio()
> indicates that we can pre-allocate bio instead of
> acquire for each time.  But I have not saw any code snippet doing such
> things in kernel. :(
> how about creating a private bio set for each o2hb_region, so that we can
> do allocation out of it?
> maybe it's faster than do allocation from global bio sets.  Also, does it
> make sense if creating a memory pool
> on each o2hb_region, so that we can have continuous pages bind to those
> bios?
>
>
> Any comments are appreciated!
>
> Thanks,
> -Jeff
>
>
>
>
>
>
>
>
> _______________________________________________
> Ocfs2-devel mailing listOcfs2-devel at oss.oracle.comhttps://oss.oracle.com/mailman/listinfo/ocfs2-devel
>
>
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120822/1b3588dd/attachment.html 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Jie Liu]

On 08/23/12 01:13, srinivas eeda wrote:
>
> On 8/22/2012 7:17 AM, Jie Liu wrote:
>> Hi All,
>>
>> These days, I am investigating an issue regarding OCFS2 unexpected
>> reboot in some real world use cases.
>> This problem occurred when the network status goes south, when the
>> disk IO load is too high, etc...
>> I suspect it might caused by ocfs2 fencing if it's BIO
>> reading/writing can not be scheduled and processed quickly, or
>> something like this happened in the network IO heartbeat thread.
>>
>> Now am trying to reproduce this problem locally.  In the meantime,
>> I'd like to ping you guys with some rough ideas
>> to improve the disk IO heartbeat to see if they are sounds reasonable
>> or not.
>>
>> Firstly, if an OCFS2 node is suffer from heavy disk IO, how about to
>> fix the bio read/write to make this IO request can not
>> be preempted by other requests? e.g,  for o2hb_issue_node_write(),
>> currently, it do bio submission with WRITE only,
>> 'submit_bio(WRITE, bio)'.   If we change the flag to WRITE_SYNC, or
>> even submit the request combine with REQ_FUA,
>> maybe could get highest priority for disk IO request.
> This was submitted before by Noboru Iwamatsu and acked by sunil and
> tao but some how didn't get merged
>
> https://oss.oracle.com/pipermail/ocfs2-devel/2011-December/008438.html
Thanks for the info!

-Jeff
>>
>> Secondly, the comments for bio allocation at o2hb_setup_one_bio()
>> indicates that we can pre-allocate bio instead of
>> acquire for each time.  But I have not saw any code snippet doing
>> such things in kernel. :(
>> how about creating a private bio set for each o2hb_region, so that we
>> can do allocation out of it?
>> maybe it's faster than do allocation from global bio sets.  Also,
>> does it make sense if creating a memory pool
>> on each o2hb_region, so that we can have continuous pages bind to
>> those bios?
>>
>>
>> Any comments are appreciated!
>>
>> Thanks,
>> -Jeff
>>
>>
>>
>>
>>
>>
>>
>>
>> _______________________________________________
>> Ocfs2-devel mailing list
>> Ocfs2-devel at oss.oracle.com
>> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>
>
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120823/86e5aeeb/attachment.html 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Jie Liu]

On 08/23/12 01:18, Sunil Mushran wrote:
> Yes. WRITE_SYNC should be good. Not FUA.
>
> Also, you may want to look into using io priorities. The code is all
> there. Just needs activation.
Yes, I'll search the list to find them out.

Thanks,
-Jeff
>
> On Wed, Aug 22, 2012 at 10:13 AM, srinivas eeda
> <srinivas.eeda at oracle.com <mailto:srinivas.eeda@oracle.com>> wrote:
>
>
>     On 8/22/2012 7:17 AM, Jie Liu wrote:
>>     Hi All,
>>
>>     These days, I am investigating an issue regarding OCFS2
>>     unexpected reboot in some real world use cases.
>>     This problem occurred when the network status goes south, when
>>     the disk IO load is too high, etc...
>>     I suspect it might caused by ocfs2 fencing if it's BIO
>>     reading/writing can not be scheduled and processed quickly, or
>>     something like this happened in the network IO heartbeat thread.
>>
>>     Now am trying to reproduce this problem locally.  In the
>>     meantime, I'd like to ping you guys with some rough ideas
>>     to improve the disk IO heartbeat to see if they are sounds
>>     reasonable or not.
>>
>>     Firstly, if an OCFS2 node is suffer from heavy disk IO, how about
>>     to fix the bio read/write to make this IO request can not
>>     be preempted by other requests? e.g,  for
>>     o2hb_issue_node_write(), currently, it do bio submission with
>>     WRITE only,
>>     'submit_bio(WRITE, bio)'.   If we change the flag to WRITE_SYNC,
>>     or even submit the request combine with REQ_FUA,
>>     maybe could get highest priority for disk IO request.
>     This was submitted before by Noboru Iwamatsu and acked by sunil
>     and tao but some how didn't get merged
>
>     https://oss.oracle.com/pipermail/ocfs2-devel/2011-December/008438.html
>>
>>     Secondly, the comments for bio allocation at o2hb_setup_one_bio()
>>     indicates that we can pre-allocate bio instead of
>>     acquire for each time.  But I have not saw any code snippet doing
>>     such things in kernel. :(
>>     how about creating a private bio set for each o2hb_region, so
>>     that we can do allocation out of it?
>>     maybe it's faster than do allocation from global bio sets.  Also,
>>     does it make sense if creating a memory pool
>>     on each o2hb_region, so that we can have continuous pages bind to
>>     those bios?
>>
>>
>>     Any comments are appreciated!
>>
>>     Thanks,
>>     -Jeff
>>
>>
>>
>>
>>
>>
>>
>>
>>     _______________________________________________
>>     Ocfs2-devel mailing list
>>     Ocfs2-devel at oss.oracle.com <mailto:Ocfs2-devel@oss.oracle.com>
>>     https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>
>     _______________________________________________
>     Ocfs2-devel mailing list
>     Ocfs2-devel at oss.oracle.com <mailto:Ocfs2-devel@oss.oracle.com>
>     https://oss.oracle.com/mailman/listinfo/ocfs2-devel
>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120823/407d3586/attachment.html 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Tao Ma]

Hi Jeff,
On 08/22/2012 10:17 PM, Jie Liu wrote:
> Hi All,
> 
> These days, I am investigating an issue regarding OCFS2 unexpected
> reboot in some real world use cases.
> This problem occurred when the network status goes south, when the disk
> IO load is too high, etc...
> I suspect it might caused by ocfs2 fencing if it's BIO reading/writing
> can not be scheduled and processed quickly, or
> something like this happened in the network IO heartbeat thread.
> 
> Now am trying to reproduce this problem locally.  In the meantime, I'd
> like to ping you guys with some rough ideas
> to improve the disk IO heartbeat to see if they are sounds reasonable or
> not.
> 
> Firstly, if an OCFS2 node is suffer from heavy disk IO, how about to fix
> the bio read/write to make this IO request can not
> be preempted by other requests? e.g,  for o2hb_issue_node_write(),
> currently, it do bio submission with WRITE only,
> 'submit_bio(WRITE, bio)'.   If we change the flag to WRITE_SYNC, or even
> submit the request combine with REQ_FUA,
> maybe could get highest priority for disk IO request.
yes, WRITE is too much easy to be preempted in cfq since it is in the
async queue, so if we have a lot of read or sync write, the heartbeat
write will be delayed a lot.

So as Srini has said in another e-mail, WRITE_SYNC should help in case
it will let you have the chance of be a sync write and be treated the
same as the read. Please check it first to see whether it is OK.

I guess the final solution will be WRITE_FUA, and I see btrfs uses it to
write out the superblock. It will be handled differently by the
underlying block layer so that it will not be in the elevator queue. It
should work but I am not sure whether we need to be so aggressive. Maybe
some test will show.

Thanks
Tao

> 
> Secondly, the comments for bio allocation at o2hb_setup_one_bio()
> indicates that we can pre-allocate bio instead of
> acquire for each time.  But I have not saw any code snippet doing such
> things in kernel. :(
> how about creating a private bio set for each o2hb_region, so that we
> can do allocation out of it?
> maybe it's faster than do allocation from global bio sets.  Also, does
> it make sense if creating a memory pool
> on each o2hb_region, so that we can have continuous pages bind to those
> bios?
> 
> 
> Any comments are appreciated!
> 
> Thanks,
> -Jeff
> 
> 
> 
> 
> 
> 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Jie Liu]

Hi Tao,

On 08/23/12 11:44, Tao Ma wrote:
> Hi Jeff,
> On 08/22/2012 10:17 PM, Jie Liu wrote:
>> Hi All,
>>
>> These days, I am investigating an issue regarding OCFS2 unexpected
>> reboot in some real world use cases.
>> This problem occurred when the network status goes south, when the disk
>> IO load is too high, etc...
>> I suspect it might caused by ocfs2 fencing if it's BIO reading/writing
>> can not be scheduled and processed quickly, or
>> something like this happened in the network IO heartbeat thread.
>>
>> Now am trying to reproduce this problem locally.  In the meantime, I'd
>> like to ping you guys with some rough ideas
>> to improve the disk IO heartbeat to see if they are sounds reasonable or
>> not.
>>
>> Firstly, if an OCFS2 node is suffer from heavy disk IO, how about to fix
>> the bio read/write to make this IO request can not
>> be preempted by other requests? e.g,  for o2hb_issue_node_write(),
>> currently, it do bio submission with WRITE only,
>> 'submit_bio(WRITE, bio)'.   If we change the flag to WRITE_SYNC, or even
>> submit the request combine with REQ_FUA,
>> maybe could get highest priority for disk IO request.
> yes, WRITE is too much easy to be preempted in cfq since it is in the
> async queue, so if we have a lot of read or sync write, the heartbeat
> write will be delayed a lot.
>
> So as Srini has said in another e-mail, WRITE_SYNC should help in case
> it will let you have the chance of be a sync write and be treated the
> same as the read. Please check it first to see whether it is OK.
We have started a test with this flag.
>
> I guess the final solution will be WRITE_FUA, and I see btrfs uses it to
> write out the superblock. It will be handled differently by the
> underlying block layer so that it will not be in the elevator queue. It
> should work but I am not sure whether we need to be so aggressive. Maybe
> some test will show.
Yes, this solution is really aggressive, the purpose is to make sure
that the bio write
will hit the disk ASAP.
I'll update the test results for you guys then.

BTW, Sunil mentioned there already has an IO priority patch set but not
yet merged.  However, I only searched
an old posts back to 2006 at:
http://www.digipedia.pl/usenet/thread/11947/7120/

Am I missing something?


Thanks,
-Jeff
>
> Thanks
> Tao
>
>> Secondly, the comments for bio allocation at o2hb_setup_one_bio()
>> indicates that we can pre-allocate bio instead of
>> acquire for each time.  But I have not saw any code snippet doing such
>> things in kernel. :(
>> how about creating a private bio set for each o2hb_region, so that we
>> can do allocation out of it?
>> maybe it's faster than do allocation from global bio sets.  Also, does
>> it make sense if creating a memory pool
>> on each o2hb_region, so that we can have continuous pages bind to those
>> bios?
>>
>>
>> Any comments are appreciated!
>>
>> Thanks,
>> -Jeff
>>
>>
>>
>>
>>
>>

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Sunil Mushran]

On Wed, Aug 22, 2012 at 9:01 PM, Jie Liu <jeff.liu@oracle.com> wrote:

> BTW, Sunil mentioned there already has an IO priority patch set but not
> yet merged.  However, I only searched
> an old posts back to 2006 at:
> http://www.digipedia.pl/usenet/thread/11947/7120/
>
> Am I missing something?
>

No, I said the code is all in there. You just need to use it.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120823/5a158b2c/attachment-0001.html 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Sunil Mushran]

On Wed, Aug 22, 2012 at 8:44 PM, Tao Ma <tm@tao.ma> wrote:

> I guess the final solution will be WRITE_FUA, and I see btrfs uses it to
> write out the superblock. It will be handled differently by the
> underlying block layer so that it will not be in the elevator queue. It
> should work but I am not sure whether we need to be so aggressive. Maybe
> some test will show.
>
>
FUA is an overkill for the heartbeat data because we don't care if that data
hits the platter or not. It can stay in the device's volatile cache.
Infact, staying
in the cache, and not hitting the platter, is better.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120823/46739e77/attachment.html 

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Tao Ma]

On 08/24/2012 01:33 AM, Sunil Mushran wrote:
> On Wed, Aug 22, 2012 at 8:44 PM, Tao Ma <tm at tao.ma <mailto:tm@tao.ma>>
> wrote:
> 
>     I guess the final solution will be WRITE_FUA, and I see btrfs uses it to
>     write out the superblock. It will be handled differently by the
>     underlying block layer so that it will not be in the elevator queue. It
>     should work but I am not sure whether we need to be so aggressive. Maybe
>     some test will show.
> 
> 
> FUA is an overkill for the heartbeat data because we don't care if that data
> hits the platter or not. It can stay in the device's volatile cache.
> Infact, staying
> in the cache, and not hitting the platter, is better.
yes, but it has the benefit that the elevator will not try to add it in
the queue and avoid any possible latency from the elevator. And my
another concern is that if ocfs2 has some high-end shared storage, it
should optimize the FUA command to be very fast and even within its
cache if the storage can survive the power outage. Having said that, it
may hurt if the backend storage doesn't implement FUA fast enough. So
maybe we can add a new parameter so that the user can decide what type
of heartbeat write he/she needs.

Thanks
Tao

[Ocfs2-devel] RFC: OCFS2 heartbeat improvements

[Jie Liu]

On 08/24/12 01:25, Sunil Mushran wrote:
> On Wed, Aug 22, 2012 at 9:01 PM, Jie Liu <jeff.liu@oracle.com
> <mailto:jeff.liu@oracle.com>> wrote:
>
>     BTW, Sunil mentioned there already has an IO priority patch set
>     but not
>     yet merged.  However, I only searched
>     an old posts back to 2006 at:
>     http://www.digipedia.pl/usenet/thread/11947/7120/
>
>     Am I missing something?
>
>
> No, I said the code is all in there. You just need to use it.
Ah, I know what I need to do.

Thanks,
-Jeff

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20120824/e2c2f110/attachment.html