Forum Discussion

Heinz_Mueller's avatar
11 years ago

SFHA 6.1: Poor performance with VxFS

Hallo @all

I'm facing poor performance with VxFS ( EMC LUNs, host-based-mirrored volume with logplex ):

We compress and uncompress a 10GB file and here are the results:

pri620[root]/database/sybase_backup {NSH}: time compress test_a

real    1m43.920s

user    1m8.303s

sys     0m14.824s

pri620[root]/database/sybase_backup {NSH}: time uncompress test_a.Z

real    3m29.777s

user    1m18.035s

sys     0m14.054s

 

Here the same but with two mirrored local disks and ZFS:

 

pri620[root]/install {NSH}: time compress test

real    1m29.672s

user    1m14.996s

sys     0m9.424s

pri620[root]/install {NSH}: time uncompress test.Z

real    1m27.838s

user    1m19.017s

sys     0m6.383s

Uncompressing is much faster!

When I create the 10GB file with mkfile I can see with iostat that

write performance is about 180-200 MB/s, but uncompressing

the file write performance is only about 55MB/s.

 

Hardware Oracle T5-2, server instance is a "Logical Domain"

I've read several documentation but I'm not able to get it faster on VxFS.

The only way it was faster is a special mount option:

mount -F vxfs -o tmplog,convosync=delay,mincache=tmpcache

But this seems to be very unsafe.

Any hints?

Regards,

Heinz

  • Hello Heinz,

    What is your volume layout?  Having mirrored array disks helps with redundancy so you can safely stripe LUNs together for better IO.  Also the /tmplog mount option is a bit risky and is not the default.  You can use other logging options but as you have noted performance will take a hit.

    If your volume is small you might need to check the block size.  Using an 8K block size will perform better for most workloads.

    I suggest taking a quick look at this article http://www.symantec.com/docs/TECH202712 .  It will help you go through the various things that affect VxFS/VxVM performance.

    I've written a tuning guide that incorporates many of the tips from support articles, admin guides, etc.  If you are interested in seeing it send me a message with your email.  Thanks.

  • Hello Heinz,

    What is your volume layout?  Having mirrored array disks helps with redundancy so you can safely stripe LUNs together for better IO.  Also the /tmplog mount option is a bit risky and is not the default.  You can use other logging options but as you have noted performance will take a hit.

    If your volume is small you might need to check the block size.  Using an 8K block size will perform better for most workloads.

    I suggest taking a quick look at this article http://www.symantec.com/docs/TECH202712 .  It will help you go through the various things that affect VxFS/VxVM performance.

    I've written a tuning guide that incorporates many of the tips from support articles, admin guides, etc.  If you are interested in seeing it send me a message with your email.  Thanks.

  • Hi

    As with most performance issue, it will likely require an investigation (Support case) to determine if VxFS has an issue. Typically VxFS performs well in multi threaded workloads.

    One notable difference is that the tests are using different devices.

    Also is the uncompress/compress function the main workload for this system ? Possibly the VxFS can be tuned to perform better for compress/uncompress, but this may affect any other workloads. 

     

    Cheers

    tony