SUMMARY: memory tuning

From: Ellen Davis <daviser_at_email.uc.edu>
Date: Mon, 18 May 1998 09:25:40 -0400

Thanks to all who replied:

"John P . Speno" <speno_at_isc.upenn.edu>
farley_at_Manassas1.TDS-GN.LMCO.COM (Patrick Farley)
kstran - Keith Strange <kstran_at_acxiom.com>
 "Allen, Mark R (PBD)" <Mark.Allen_at_pbdir.com>
Kurt Carlson <snkac_at_java.sois.alaska.edu>
Willig Reimund <Willig.Reimund_at_gdr.de>
alan_at_nabeth.cxo.dec.com

Many people who replied suggested that I simply needed more RAM and swap
space. Alan from DEC suggested that there are other considerations. (See
his message below.) I have decided to collect system statistics in order
to make a better determination of the problem and the best solution.

Many people suggested various including:

top
monitor gatekeeper.dec.com:/pub/DEC/monitor.alpha.tar.Z
collect also from gatekeeper (collects data and then displays it using a
GUI interface)
There is a built in program in CDE under system administration then click
on tuning then process tuning. From there you can configure it to show
real/vm per process in descending order.

Randy Hayman has sevaral tools available including:
iostat like tool:
        ftp://raven.alaska.edu/pub/sois/uaio-v2.0b.tar.Z
ps like tool:
        ftp://raven.alaska.edu/pub/sois/uakpacct-v2.0.tar.Z
vm_mon, syd, disks performance monitoring tools:
ftp://raven.alaska.edu/pub/randy/perf_mon_tools/



alan_at_nabeth.cxo.dec.com:
        Measuring free memory on modern UNIX systems is hard, because
        most of them support a unified buffer cache (UBC). Older versions
        set aside a fixed percentage of memory for the file system
        buffer cache and left everything else for processes. Today,
        the buffer cache and processes share the same available
        memory; unified.

        Before, a low memory condition meant memory had to be taken
        away from processes, which nearly always meant page-outs.
        On UBC systems the first quick source for memory is cache
        buffers that haven't been used recently and don't hold
        dirty (writing) data. If there aren't any of those, then
        dirty buffers and program data are probably a toss up,
        with dirty buffers having a slight edge since they need
        to be written anyway.

        So, a system with lots of I/O going on, may appear to always
        be "low" on memory, simply because the buffer cache has most
        of it. Where you get performance problems is when you have
        normal processes wanting memory and programs doing lots of
        I/O that use large amounts of the cache. In such a case,
        you can limit the amount of memory that the cache will use.

        One of the VM subsystem configuration parameters is the amount
        of free memory that the system considers "low". When it get
        to this point, the system will actively start looking for
        ways to get memory back; invalidate cache buffers, write
        dirty buffers, write process pages out, etc. I think this
        number is 128 pages by default, but I haven't paid close
        attention to such things in a long time.

        Knowing what the threshold values are between what's acceptable
        and what isn't is hard, because it depends on the applications.
        Background processing may be able to with stand high levels
        of paging then interactive use that wants quick response. One
        thing you can look at is to consider that each page-out operation
        requires an I/O request. A high rate of page-out also means
        lots of I/Os which may saturate disks. To find the upper
        bound for saturation I/O, do sequential page size reads to
        your page/swap device. That gives you a best case for how
        much I/O a given disk can handle. Then, when the system is
        paging or slow see where the I/O is going. If it is only the
        paging device(s) and the pageout rate is above normal, then
        more memory could help. If the system is slow and there is
        I/O on other devices, then it may be the I/O that's the problem.


 Ellen Davis
 Ellen.Davis_at_uc.edu
Received on Mon May 18 1998 - 15:29:21 NZST

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:37 NZDT