performance issues with largefiles enabled?

[Christine Hartman]

I was wondering do you really take a performance hit if you enable largefiles on your filesystem? Let's say you have a filesystem that has largefiles enables and it contains hundreds of thousands of files...some over over 2GB and many are not. Will I see a performace impact or waste disk space? Does the inode table get impacted?

[A. Clay Stephenson]

The performance hit comes not from largefiles but rather from the large number of files. At the very least, directory searches are going to be expensive. Operations which alter directories are also going to be expensive.
The inode table is dynamic and will not be a problem - if this is a vxfs filesystem (and it should be). In general, whenever you need hundreds of thousands of files in one filesystem, you are probably not thinking of the problem in the optimum way. You are essentially using a filesystem as a database -- something it was really never designed to be. Because vxfs is extent based, mixing small and large files will not be wasteful of space.

[S.K. Chan · Accepted solution]

The answer is no, no performance impact whatsoever and enabling largefiles option has nothing to do with inode table at all. Just be aware that backup programs/utilities like tar and cpio are not able to handle files larger than 2GB (I'm sure you know that already). Having massive number of files (regardless of its size) in a single directory would impact performance, still it has nothing to do with whether largefiles option is enabled or not.

[Christine Hartman]

it is vxfs. Just wanted to inquire, because if I make a filesystem with largefiles enabled and the DBA decides to use that filesystem for his oracle archive logs..I want to make sure it's not going to have a performance impact on that Oracle functionality. And in our documentum environment we will have hundreds/thousands of files of various sizes (i've limited each filesystem to be no more than 100GB...to ease sys admin duties a little), also wanted to make sure that if largefiles is enabled there won't be a performance hit or wasted disk space "because" of largefile being enabled.

Thanks.

Christine

[James Murtagh]

Hi Christine,

Its a very good question, although I think there are actually quite a few questions in your post, most of which have been addressed already. The only other point I will make is that there is a _slight_ performance penalty involved with having a file larger than two gigabytes in size, though this does not mean that enabling largefiles has a performance impact! The 32 bit binaries hpux uses need to use some sort of indirection to manage the amount of extents required for a "large file". This will cause a very slight slowdown in performance, but this will not be noticeable. I haven't looked into the forthcoming tru64 advFS in detail but I'd imagine it uses a true 64 bit (no pun intended) representation. Also note that from HPUX 11.23 largefiles will be the default config, so that in itself implies there is no real performance impace.

Obviously what is slow to a computer will not be slow to us, think of when you enter input at a (virtual) terminal - you think it responds instantaneously but it takes an eon in cpu terms.

Regards,

James.

[Con O'Kelly]

Hi Christine

I think the others have answered your question, however I would make one point as you seem to be looking at the performance of Oracle filesystems.

If you are setting up Oracle filesystems, I would keep your archive logs & data files in seperate filesystems. You may already be doing this & may be aware that if you have Online JFS you can use the "mincache=direct,convosync=direct,delaylog"mount options for the filesystems containing your datafiles.
In some cases this will improve the performance.
Its something you may want to look into. There's been quite a bit of discussion about this in the past on the forums.

Cheers
Con

[Christine Hartman]

yes thanks. We do isolate Oracle filesystems from other data, although we leave up to our dba where he puts his dbf files and archive logs. He does usually keep them separated too. We will be using online jfs (and vxvm) on our production servers so I will keep those options in mind.

Thanks for all of the feedback!