[Linux-cluster] GFS performance.

[Vikash Khatuwala]

Hi,

Can I downgrade the lock manage from lock_dlm to no_lock? Do I need 
to un-mount the gfs partition before changing? I want to see if it 
makes any performance improvements.

Thanks,
Vikash.


At 11:18 AM 21-04-09, Vikash Khatuwala wrote:
>Hi,
>
>I am using Virtuozzo OS visualization which does not have a single 
>file for the entire VM's filesystem. All VMs are simply 
>sub-directories and OS files are stored in a common templates 
>directory which is sym linked to back to the VM's directory, so if 
>an OS file is changed inside the VM then the symlink breaks and a 
>new file is put in the VM's private directory. I cant use GFS2 
>because it is not supported by Virtuozzo. All VMs are simply running 
>web/db/ftp.
>
>So this basically means that there are a lot of symbolic links 
>(small files). The GFS has a block size of 4K so I also chose 4K as 
>my block size for my performance testing to asses the worst case 
>scenario. If I change the block size to 256K then the performance 
>difference between ext3 and GFS are minimal. Also when I migrate the 
>VM out from GFS(RAID5 SAS 15K) to ext3(single disk SATA), there is a 
>significant noticeable performance gain!
>
>Below tests are on the same disk set (5 disk RAID5 SAS 15K) with 2 
>partitions, GFS and ext3.
>Results at 4K random reads:
>GFS : about 1500K/s
>ext3 : about 7000K/s
>
>Results at 256K random reads:
>GFS : about 45000K/s
>ext3 : about 50000K/s
>
>Results at 256K sequential reads:
>GFS : over 110,000K/s (my single GB NIC maxes out)
>ext3 : over 110,000K/s (my single GB NIC maxes out)
>
>fio test file as below only rw and blocksize were changed for the 3 
>different scenarios above.
>[random-read1]
>rw=randread
>size=10240m
>directory=/vz/tmp
>ioengine=libaio
>iodepth=16
>direct=1
>invalidate=1
>blocksize=4k
>
>[random-read2]
>rw=randread
>size=10240m
>directory=/vz/tmp
>ioengine=libaio
>iodepth=16
>direct=1
>invalidate=1
>blocksize=4k
>
>Thanks,
>Vikash.
>
>
>At 01:00 AM 21-04-09, Jeff Sturm wrote:
>> > -----Original Message-----
>> > From: linux-cluster-bounces at redhat.com
>> > [mailto:linux-cluster-bounces at redhat.com] On Behalf Of Vikash
>> > Khatuwala
>> > Sent: Monday, April 20, 2009 11:23 AM
>> > To: linux-cluster at redhat.com
>> > Subject: [Linux-cluster] GFS performance.
>> >
>> > OS : CentOS 5.2
>> > FS : GFS
>>
>>Can you easily install CentOS 5.3 and GFS2?  GFS2 claims to have some
>>performance improvements over GFS1.
>>
>> > Now I need to make a decision to go with GFS or not, clearly
>> > at 4 times less performance we cannot afford it, also it
>> > doesn't sound right so would like to find out whats wrong.
>>
>>Be careful with benchmarks, as they often do not give you a good
>>indication of real-world performance.
>>
>>Are you more concerned with latency or throughput?  Any single read will
>>almost certainly take longer to complete over GFS than EXT3.  There's
>>simply more overhead involved with any cluster filesystem.  However,
>>that's not to say you're limited as to how many reads you can execute in
>>parallel.  So the overall number of reads you can perform in a given
>>time interval may not be 4x at all (are you running a parallel
>>benchmark?)
>>
>>Jeff
>>
>>
>>--
>>Linux-cluster mailing list
>>Linux-cluster at redhat.com
>>https://www.redhat.com/mailman/listinfo/linux-cluster