VxFS versus hfs

[Mott Given]

Does anyone have any numbers that quantify the performance benefit of converting hfs filesystems to VxFS ones?

Has HP or anyone else done any benchmarks on this that they would share with us?

I am asking this in regard to HP-UX v11.0 systems running Oracle database applications.

[Tom Maloy]

Don't know about operating performance, but VxFS will come up much quicker after a "hard" shutdown if logging is enabled...

Worth the conversion just for that.

Tom

[harry d brown jr]

Are you asking in regard to the Oracle database on an VxFS filesystem versus an HFS?

System recovery is faster with VxFS, and I think HP plans on phasing out HFS anyways. The "overhead" of VxFS" is minute and isn't worth measuring.

live free or die
harry

[Martin Johnson]

With OnlineJFS you can extend VxFS file systems without having to umount them.

HTH
Marty

[Mott Given]

To clarify my original question, I was not concerned about the "overhead" of VxFS - I was wondering about its performance benefits over hfs.

If I have a Oracle system with some I/O performance problems, will this conversion from hfs to VxFS offer significant performance gains or just 1 or 2% gains?

[S.K. Chan]

Numbers ? I don't think so because hfs is better when you got a lot of small files that gets updated frequently. So hfs and vxfs are never being compared in that manner (quantitatively). Having said that the benefits of vxfs outweigh hfs significantly.
Take a look at the general comparison between vxfs and hfs below ..
DocID=KBRC00008849
http://support2.itrc.hp.com/service/cki/docDisplay.do?docLocale=en_US&docId=200000058826912
Hope it helps ..

[keith persons]

Don't have any immediate numbers available, but the last time HFS worked 'better' than VxFS was hp-ux 10.10 - had JFS v2.0 and did have a reputation of being slow. With 10.20 and above JFS v3.0 or or higher is used and there are a couple of advantages - first is boot time under normal circumstances - much faster w/JFS. Block sizes are also more dynamic (depending on FS size) and kernel table management for JFS inodes completely dynamic (unlike ninode w/HFS).

There's also the pending obsolescence of HFS and with Online JFS there are mount options that can be advantageous for Oracle.

Should I discover some measurements that would be useful I will post them here.

Keith

[A. Clay Stephenson]

I'm not going to go search for performance differences; I assume that you can do that as well as I. However, I will relate a true story that occurred while I was taking the 'Managing Oracle on HP-UX' course - this was a class taught jointly by Oracle and HP but it's no longer available. The Oracle guy reccomended HFS for better performance and I politely said 'Bull----'. We them proceeded to setup an equivalent test environment and in that test, VxFS was about 4-5% faster.

After many years using both, I have not seen a single case within the last 5 years or so that HFS outperformed VxFS unless asynch io was used.

[Sean OB_1]

In my experience you won't see any significant I/O improvements in moving from HFS to vxfs.

vxfs does have a small system overhead due to the journalling.

However you'll see an enormous difference in boot times between hfs and vxfs.

VxFS comes up much, much faster, and is worth the coversion for that alone, in my opinion.

Sean

[Patrick Wallek]

If you are having I/O problems, switching from HFS to VxFS probably wouldn't buy you much more than Clays 4-5%. What you need to look at instead is how your VG's and LV's are set up. Is everything RAID5, or RAID 1 or RAID0?

Your VG and LV layout, striping, stripe size, etc., will have a large impact on I/O performance.

[A. Clay Stephenson · Accepted solution]

One point cannot be overstressed. The place to start getting performance out of Oracle is almost always in the SQL itself. If you are seeing large amounts of IO (and that's generally large amounts of sequential IO) there is really only one fix - the SQL code. In the vast majority of cases you can tune the OS and Oracle parameters till the cows come home and you might double the performance (20% is much more common) but if you need a 5x-10x increase in performance there's really only 1 place to get it - the code itself. Often the addition of a single index can offer an order of magnitude increase in throughput.

[James R. Ferguson]

Hi Mott:

Performance aside, VxFS filesystems offer not only faster recovery (as noted); the ability to dynamically resize filesystems (with the Online JFS license); unlimited inode growth; the ability to tune performance on a filesystem basis by utilizing different mount options (see the man pages for 'mount_vxfs'); and the ability to create snapshots for "online" backups.

Based on the above, alone, I think VxFS filesystems are far superior to HFS ones -- performance aside.

Regards!

...JRF...

[Shannon Petry]

I'll give you another side to the VxFS and HFS scenario.
While I tune and maintain some oracle servers my bread and butter is FEA (Finite Element Analysis) or computer simulation code.

Having been in the market since the mainframe days, I have seen and tested every new technology you can think of. So...

Pure I/O VxFS is faster than HFS in my testing. While only a few percentage points which is based on file sizes, it does make a difference.
Now you have to tune your FS for what your doing, and make sure that your covering as many spindles as possible. Normally my FEA scratch disks stripe size is max size, and block size is huge (sorry not giving away trade secrets). Most of the scratch files the FEA solvers write are between 1.5 and 20GB, so nothing small.
HFS may be faster for small files, but Im not so sure.

Lastly, as mentioned before by others VxFS can recover faster. This also translates into less data lost in case of a outage.
I used to loose alot of data in the HFS days, as FEA code tends to cause panic/dump on hard runs with bad data. Now data loss is almost a non issue. (I still run backups though).

Regards,
Shannon