Re: Very slow samba file transfer speed... any ideas ?

[Martin Steigerwald <Martin@lichtvoll.de>]

Am Freitag, 20. Juli 2012 schrieb Shavi N:

> On Fri, Jul 20, 2012 at 12:19 AM, Martin Steigerwald
> 
> <Martin@lichtvoll.de> wrote:
> > Am Donnerstag, 19. Juli 2012 schrieb Shavi N:
> >> Hi,
> > 
> > Hi Shavi,
> > 
> >> Thanks.
> >> 
> >> This is the output:
> >> btrfs:
> >> $ dd if=/dev/zero of=/mnt/shared/misc/temp_file bs=1M count=1400
> >> 1400+0 records in
> >> 1400+0 records out
> >> 1468006400 bytes (1.5 GB) copied, 1.56841 s, 936 MB/s
> >> 
> >> ext4:
> >> $ dd if=/dev/zero
> >> of=/mnt/500/VirutalBox_VMs/thor/thor-data/temp_file bs=1M
> >> count=1400
> >> 1400+0 records in
> >> 1400+0 records out
> >> 1468006400 bytes (1.5 GB) copied, 4.98993 s, 294 MB/s
> > 
> > Did you actually read and understand my mail?
> > 
> > Do you really think that your local disks is/are yielding that
> > transferrate? (Unless you are using BTRFS RAID 0/10 on lots of
> > disks.)

> Yes I read and understood your email. my btrfs volume consists of 11
> HDD's. I was very surprised with that result myself...

I think that you did not understood it. Why? Your dd tests where without 
conv=fsync – did you even try with conv=fsync to compare results?

11 really fast 15000rpm FC / SAS disks could possibly do 936 MB/s. But 
regular 7200rpm SATA disks depending to the on disk location might be as 
slow as 40-50 MB/s – just try fio disk-zone-profile on one if you do not 
believe this – and then even with 11 disks 936 MB/s is out of reach.

And then speed depends on the BTRFS RAID level as well. And if BTRFS is 
using compression then testing with zeros is bogus anyway.

Also its questionable whether CIFS will use 1 MiB blocksizes. It might be 
using it in most current kernels, since there have been some adaption  – 
but I am not sure ATM whether they have been for reads or writes, they 
have not been for both, check wsize, rsize or similar named option 
maximums –, but thats also a point where to look.

But then are the files that you transfer there that big at all? And what 
is the accesses pattern? Virtualbox VM images are usually not written 
sequentially to, so you need a random I/O workload.

I do think in order to get more close to the possible reason of Samba 
slowness with BTRFS a simple dd test won´t help one bit. Even the 
conv=fsync case will most likely not be testing the workload that Samba 
exercises on the local disks.

Some fio random I/O workload or dbench or filebench workload might be much 
closer.

I think currently you are measuring something that has almost nothing to 
do with your real workload.

Ciao,
-- 
Martin 'Helios' Steigerwald - http://www.Lichtvoll.de
GPG: 03B0 0D6C 0040 0710 4AFA  B82F 991B EAAC A599 84C7