On 10 Jan 00, at 21:04, Thomas Leavitt wrote:
- The one problem I do have, is that during a high sequential write
load, like creating an Oracle tablespace, the high sequential write rate locks out other processes from doing I/o to the netapp. For example, during a create of a 2gb db file, If I (in another window) try and do anything that does I/O, I can wait for 5/10/15 seconds for the command to come back. I'm not sure why this is, but I think this is a problem in AIX, NOT the netapp - because if during the massive sequential writes I try and access the netapp from another computer I get great response. AIX must be queing NFS I/O's or something - which doesn't make sense!
Yes, NFS version 3 does client side caching. You could try switching to NFS version 2, but that might hurt your overall performance.
In my experience with AIX, it likes to buffer up a lot of writes in RAM, and when RAM is exhausted, AIX has to flush out those writes to the NFS server. While AIX is busy flushing the buffers, its interactive response can get very sluggish, just as you have seen.
I have seen the following behavior when copying a very large file from a local AIX disk to a netapp.
At first the disk reads on AIX go way up, but the network load stays low. On the netapp, the network, CPU and NFS ops stay low. Then suddenly AIX stops reading the disk and the network load goes way up. This drives up the network, NFS, and CPU load on the netapp. Then AIX goes back to reading the disk and the network load drops to nothing, etc.
It looks like AIX if flip-flopping between
1) reading the disk and buffering NFS writes to RAM, and
2) Flushing the buffered NFS writes to the netapp.
It's not doing both jobs simultaneously. This actually gets more pronounced the more RAM you have. Apparently AIX is more than happy to exhaust ALL of its RAM before commencing to flush the writes. And when you've got 1G or 2G of RAM, that's a lot to flush.
I think there are some tuning parameters that you can fiddle with to get AIX to be less bursty. You can tell AIX to go ahead and start flushing RAM buffers sooner. I don't know if this would help overall throughput or not. I haven't played with this at all, but our resident AIX expert has.
Steve Losen scl@virginia.edu phone: 804-924-0640
University of Virginia ITC Unix Support