This was solved using dxkernaltuner. Thanks to Thomas Blinn, Alan Rollow,
Yogesh Bhanu, Thomas Sjolshagen, Alan Davis, Kevin Raubenolt and Nicola
Milutiniovic (I think that's all, sorry if anybody is missing), for many
suggestions as to which parameters to tune and how.
The old config that won't run:
> ulimit -a
time(seconds) unlimited
file(blocks) unlimited
data(kbytes) 131072
stack(kbytes) 8192
memory(kbytes) 243792
coredump(blocks) unlimited
nofiles(descriptors) 4096
vmemory(kbytes) 4194304
the new config that does run:
> ulimit -a
time(seconds) unlimited
file(blocks) unlimited
data(kbytes) 248260
stack(kbytes) 8192
memory(kbytes) 243896
coredump(blocks) unlimited
nofiles(descriptors) 4096
vmemory(kbytes) 4194304
The crucial kernal variable seems to be
per_proc_data_size = 254217728
which was previously:
per_proc_data_size = 134217728
Also max_per_proc_data_size was tuned to be the difference between
max_per_proc_address_space and max_per_proc_stack_size, ubc.maxpercent was
reduced to 50%, gh_chuks was turned on, as was vm_bigp_enabled, maxusers
was reduced from 256 to 128, vm_swap_eager was turned off, and
max_proc_per_user was increased to 512. Some of these may be non-essential
to solving the problem, but the machine seems to be happy.
Thanks again to all
Derek
_________________________
Derek Gatherer Ph.D. Cert.Ed.
Computer Officer
Institute of Virology
Church St.
Glasgow G11 5JR
Phone: +44 141 330 6268
Fax: +44 141 337 2236
Email: d.gatherer_at_vir.gla.ac.uk
Website:
http://www.vir.gla.ac.uk
Received on Mon Feb 23 2004 - 09:37:40 NZDT