[OpenAFS] AFS namei file servers, SAN, any issues elsewhere? We've had some. Can AFS _cause_ SAN issues?

Jason Edgecombe jason@rampaginggeek.com
Thu, 20 Mar 2008 18:01:10 -0400

Yes. We only have one fibre channel HBA and one fibre channel disk pack, 
a Sun StorEdge 3511 expansion tray with SATA disks.

For what it's worth, we just tested 1.4.6 inode fileserver (nologging 
ufs) on an old-style direct-attached SCSI disk pack and saw similar 
sluggish vos performance to what we saw on our SAN disk pack running the 
1.4.1-inode fileserver.


Kim Kimball wrote:
> Thanks, Jason.
> Is the hardware the same as what you tested last year?
> Kim
> Jason Edgecombe wrote:
>> Is this what you need?
>>   PKGINST:  SUNWsan
>>      NAME:  SAN Foundation Kit
>>  CATEGORY:  system
>>      ARCH:  sparc
>>   VERSION:  1.0
>>   BASEDIR:  /
>>    VENDOR:  Sun Microsystems, Inc.
>>      DESC:  This package provides a support for the SAN Foundation Kit.
>>    PSTAMP:  sanserve-a20031029172438
>>  INSTDATE:  Jan 15 2008 10:37
>>   HOTLINE:  Please contact your local service provider
>>    STATUS:  completely installed
>>     FILES:       22 installed pathnames
>>                   4 shared pathnames
>>                   1 linked files
>>                  11 directories
>>                   2 executables
>>                 239 blocks used (approx)
>> Running Solaris 9 09/05HW Sparc with Sun SAN foundation.
>> Jason
>> Kim Kimball wrote:
>>> Hi Jason,
>>> Thanks!
>>> Can you tell me which flavor of SAN you're using?
>>> Kim
>>> Jason Edgecombe wrote:
>>>> Robert Banz wrote:
>>>>> AFS can't really cause "san issues" in that it's just another 
>>>>> application using your filesystem.  In some cases, it can be quite 
>>>>> a heavy user of such, but since its only interacting through the 
>>>>> fs, its not going to know anything about your underlying storage 
>>>>> fabric, or have any way of targeting it for any more badness than 
>>>>> any other filesystem user.
>>>>> One of the big differences that would effect the filesystem IO 
>>>>> load that occurred between 1.4.1 & 1.4.6 was the removal functions 
>>>>> that made copious fsync operations.  These operations were called 
>>>>> in fileserver/volserver functions that modified various in-volume 
>>>>> structures, specifically file creations and deletions, and would 
>>>>> lead to rather underwhelming performance when doing vos restores, 
>>>>> deleting, or copying large file trees.  In many configurations, 
>>>>> this causes the OS to pass on a call to the underlying storage to 
>>>>> verify that all changes written have been written to *disk*, 
>>>>> causing the storage controller to flush its write cache.  Since 
>>>>> this defeats many of the benefits (wrt I/O scheduling) on your 
>>>>> storage hardware of having a cache, this could lead to overloaded 
>>>>> storage.
>>>>> Some storage devices have the option to ignore these calls from 
>>>>> devices, assuming your write cache is reliable.
>>>>> Under UFS, I would suggest that you'd be running in 'logging' mode 
>>>>> when using the namei fileserver on Solaris, as yes, fsck is rather 
>>>>> horrible to run.  Performance on reasonably recent versions of ZFS 
>>>>> were quite acceptable as well.
>>>> I can confirm Robert's observations. I recently tested openafs 
>>>> 1.4.1 inode vs 1.4.6 namei on solaris 9 sparc with a Sun Storedge 
>>>> 3511 Expansion tray fibre channel device. The difference is 
>>>> stagerring with vos move and such. We have been using the 1.4.6 
>>>> namei config on a SAN for a few months now with no issues.
>>>> Jason