On Thu, 18 Feb 1999, Todd C. Merrill wrote:
with occasional (every 5-10 minutes) sustained peaks of:
77% 1242 1201 0 489 1731 5285 5049 0 0 6 81% 1497 1279 0 600 2048 4327 5026 0 0 5 78% 1495 1221 0 563 1589 4287 5488 0 0 4 76% 1229 1213 0 472 1795 4546 5262 0 0 3 81% 1467 1150 0 1345 2634 3683 5746 0 0 2 80% 969 1068 0 397 2051 3876 6298 0 0 2 91% 1049 905 0 384 1102 1230 10622 0 0 2 92% 1328 942 0 448 1801 1536 10336 0 0 2
The inbound network traffic (less than 1MB/sec) doesn't jive withyour disk writes (5-10MB/sec).
This suggests some sort of meta-data updates, e.g., 'touch *' though in that example inode locality should be pretty good and thus you'd not see too many blocks written per operation.
Hmmm ... I just noticed that you've got CIFS as well as NFS. Some MicroSoft applications (e.g., Word or Excel) treat their files like databases. If you didn't have any CIFS client write caching (due to op locks being disabled or revoked on account of multiple accesses to the file) then saving one of these files might product this sort of pattern.
It would be helpful to try to isolate whether NFS or CIFS is causing this pattern. If NFS, do an nfsstat at the beginning and end of one of these fits then compare them to see what's being done. If CIFS is the culprit, I'll have to defer further diagnosis to someone else as it's been several years since I looked at our CIFS instrumentation very closely.
-- Karl