Very interesting System Behaviour

From: Sreekumaran Padiyath <Kumar.Padiyath_at_psi.ch>
Date: Thu, 25 Mar 1999 15:02:13 +0100

  
  Hi,
   One of our best tru64 user has noticed the following behaviour of one of our Alpha4100(having 3 processor). If any one can help us
I will be very thankful.

The following is the present configuration of proc and vm:
=========================================================

proc:
        maxusers = 256
        max-proc-per-user = 256
        max-per-proc-address-space = 5368709120
        per-proc-address-space = 5368709120
        max-per-proc-data-size = 5368709120
        per-proc-data-size = 5368709120
        max-per-proc-stack-size = 1048576
        per-proc-stack-size = 1048576

vm:

        vm-vpagemax=33554432
        vm-maxvas = 5368709120
        ubc-minpercent = 1
        ubc-maxpercent = 3
        ubc-borrowpercent = 2
        new-wire-method = 0
        vm-aggressive-swap = 0
        vm-page-free-target = 1024
        vm-page-free-swap = 296
        vm-page-free-hardswap = 2048
        vm-page-free-min = 80
        vm-page-free-reserved = 80
        vm-page-free-optimal = 296
        vm-page-prewrite-target = 2048


Problem: System locks up when running programs with large memory needs
============================================================ ========= =
My system is a Alpha server 4100, presently running with 3 400MHz processors
and 2GByte of RAM under Digital Unix 4.0D.
When running programs whose RAM needs are below about 1.8 GByte, there is
no problem with the system. However, when the RAM needs of a program exceed
about 1.8-1.9 GByte the system gets essentially locked up: no logins are
possible and previously running foreground jobs do not respond (top or iostat
e.g. cease to print out system activities).

I made sure that in my programs which produce the above problem, not all of
the allocated RAM is active at once, indeed in the code which produces these
problems, only 2 out of 3 very long vectors are used simultaneously.
Indeed, I can avoid the memory-management bottleneck by moving one of the three
vectors to disk and freeing the corresponding memory.
When the system locks up, it stays locked for many hours, but after a long
time (10-20 hours) it 'heals'. While locked, there is no disk activity
no swapping or pageing.

On a Linux-Intel system I have not been able to reproduce the problem
(no locking-up occurs).

The problem has been reported to Digital/Compaq repeatedly. Some system
parameters have been modified without success.

Suggestions are welcome!
============================================================ =======

   Thanks in advance

Please send the mail in the following address. I will summarize.

PADIYATH Sreekumar | System Manager | Paul Scherrer Institute
CH-5232 Villigen PSI | Switzerland | Tel: +41 56 310 36 43
FAX : +41 56 310 36 49 | E-Mail: kumar.padiyath_at_psi.ch
Received on Thu Mar 25 1999 - 14:04:52 NZST

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:39 NZDT