[K42-discussion] control flags for ReAIM benchmark

[Bryan S Rosenburg]

Andrew Baumann wrote on 04/09/2006 10:27:42 PM:
> Hi all,
> 
> I'm trying to determine the "best" set of kernel control flags to use 
for the 
> ReAIM benchmark. The workload is quite configurable, but generally it 
seems 
> to be a fork-intensive benchmark intended to measure system throughput 
for a 
> multiuser workload, doing various kinds of IPC and operations on small 
files.
> 
> I've experimentally determined that control flags 0x01ef seem to return 
the 
> best results. This corresponds to:
> #    bit  0 [0x0000000000000001] (RUN_SILENT)         ON
> #    bit  1 [0x0000000000000002] (DISABLE_IO_CPU_MIGRATION)   ON
> #    bit  2 [0x0000000000000004] (SLOW_THINWIRE_POLLING)   ON 
> #    bit  3 [0x0000000000000008] (UID_PROCESSOR_ASSIGNMENT)   ON
> #    bit  4 [0x0000000000000010] (DYN_PROCESSOR_ASSIGNMENT)   OFF
> #    bit  5 [0x0000000000000020] (PAGING_OFF)         ON
> #    bit  6 [0x0000000000000040] (NON_SHARING_FILE_OPT)      ON
> #    bit  7 [0x0000000000000080] (USE_MULTI_REP_FCMS)      ON
> #    bit  8 [0x0000000000000100] (NO_ALLOC_SANITY_CHECK)   ON
> #    bit  9 [0x0000000000000200] (SMALL_FILE_OPT)      OFF
> #    bit 10 [0x0000000000000400] (DBG_FLAG)         OFF
> #    bit 11 [0x0000000000000800] (SLOW_EXEC)         OFF
> 
> ... since I'm not familiar with the exact meaning of these flags, I 
wanted to 
> check whether that configuration seems sane. Can anyone explain what the 

> SMALL_FILE_OPT and UID/DYN_PROCESSOR_ASSIGNMENT flags do?

The UID/DYN_PROCESSOR_ASSIGNMENT flags don't do anything.  There's no use 
at all of the "DYN" flag, and the only uses of the "UID" flag are 
commented or ifdefed out.

Dilma will have to remind us of what the SMALL_FILE_OPT flag is good for.

For SDET, we boot the machine with flags 0xfff but then switch to 0xf7f 
for the actual experiment.  The result is that the ram filesystem is 
populated entirely with multi-rep FCMs.  In particular, the executable 
files that are shared by all the scripts are backed by low-contention 
objects.  The temporary files created by a particular script, however, use 
single-rep FCMs, which have lower creation/destruction overhead and have 
no down side if they're not shared.

- Bryan

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://ozlabs.org/pipermail/k42-discussion/attachments/20060410/067a0316/attachment.htm