I just changed the scheduler from the default cfq (which is said to have the best multiuser performance) to anticipatory and then deadline. Wow, what a difference. Anticipatory takes about 10 seconds to list a directory of 5,000 items under load from dd, and deadline takes about 5 (yes, i dropped the caches and buffers first).
Is this normal behavior with cfq? I was under the assumption that it would be most ideal for this type of workload. On Nov 8, 2007 11:42 PM, Ryan Bair <[EMAIL PROTECTED]> wrote: > I have Debian Etch AMD64 installed on a Dell PowerEdge 1950, connected > to a MD1000 (SAS/SATA array) via a PERC5/e SAS controller. The array > is RAID5. The whole drive is LVM, 9TB are one big XFS partition. > > > Linear reads/writes are very fast, directory listings are also fast. > However, if the disk is under any kind of load (say someone > downloading a file at 20MB/s) long directory listings (ls -l) get > extremely slow. Listing a directory of 1000 files can take 5 minutes. > Normal directory listings remain reasonable. > > If I strace ls -l, I notice its stopping at each getxattr() call for a > split second, seems that's where it is wasting all of its time. > > Upon a recommendation at #xfs, I tried using attr2 via the mount > option. xfs_db -c version shows: > versionnum [0xb094+0x8] = V4,ATTR,ALIGN,DIRV2,EXTFLG,MOREBITS,ATTR2 > > The whole install is pretty vanilla. Any thoughts on why this system > can't handle this type of load? Please cc me as I am not subscribed to > this list. > > Thanks > -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]