Thanks so much for the answers to my question. I had at least 17 responses.
Almost all of you suggested that the Advanced File System was the problem.
I do have AdvFS file systems for users, mail, local programs and source,
tmp, and var. We will be changing all but users back to UFS soon.
As most of you already knew, AdvFS "funnels" through CPU 0 currently, UFS
is truly SMP as of 3.0, and AdvFS will be SMP as of 4.0.
Here is my original post:
>We have an Alpha 2100 4/275 with 2 processors and 1 gigabyte of memory. We
>have over 14,000 accounts and between 100 and 250 users typically logged in.
>We are currently running 3.2D-1.
>We have noticed that most of the kernel "system" overhead seems to only happen
>on CPU 0, as if DU operated as a master/slave system. Recently, our "system"
>time on CPU 0 has been running 80-90%, causing load averages to get as
>high as 40 or more occasionally. CPU 1 appears to have plenty of extra
>cycles during these periods and is being used by 1 or 2 long running jobs.
>If we stop the long running (CPU bound) jobs, this still does not free up
>CPU 1 to help out with the system load on CPU 0.
>Is DU really a master/slave system (I/O and possibly certain other kernel
>processing only happening on CPU 0)? If so, will DU 4.0 fix this?
Paul E. Rockwell from Digital Equipment Corporation at Rocky Hill, CT
(rockwell_at_rch.dec.com) gave me this answer:
>...Most kernel subsystems are parallelized, however there are
>a few exceptions. Most notable is AdvFS. It is "funnelled" in versions of
>Digital UNIX before 4.0 - this means that all AdvFS activity is handled by
>the boot processor (4.0 makes AdvFS parallelized). UFS code, however, is
>parallelized in all releases past 3.0.
Again, thanks to all those who answered.
Greg Bell
Received on Fri Mar 01 1996 - 00:30:01 NZDT