Hi,

I'd tend to suggest first sticking close to the benchmarks that directly 
measure OS performance and various services like: IO tests, Network Tests, and 
Micro-benchmarks ( ex: libmicro).  Getting this lower level tests well done 
would help everyone and haven't been done well before.  And it has the benefit 
of directly telling you what to improve.

The problem with many of the application benchmarks that have been suggested is 
that often the are very susceptible to application tuning and compiler tuning 
differences.  It also would take a LOT of work to make sure the apps tests are 
testing what you think they are.  SPEC and TPC and ISVs (that have benchmarks) 
spend huge amounts of time trying to address these issues.

Someone suggested Bonnie, but it still has some issues...
A nearly single threaded test optimized for single disk performance, it does 
not work well with modern striped or RAID disks. bonnie++. Has some 
improvements over bonnie: can disable putc/getc with "-f" flag.

Some other thoughts on benchmarks...
Postmark
A file system benchmark developed by Network Appliance and reputed to emulate 
the operation patterns of email users and netnews servers. Primary liability is 
the question of disk vs CPU benchmarking: since postmark tests very small 
transactions, it can be turned into a fairly heavy CPU benchmark though it 
claims to be testing IO. As a CPU benchmark, it suffers from being single 
threaded. As a file system benchmark, it offers limited testing of file system 
operations (only create/delete and read/append), and operation mix and size is 
not standardized and thus can be tuned to the system under test. No central 
publication authority or peer review exists to ensure fair and accurate usage 
of the benchmark.

As a corollary to mpogue's "Most simple benchmarks suck", I'd add
"Most complicated benchmarks don't measure what you assume they do"

-- Brad (I'm a sun performance/benchmark guy)
This message posted from opensolaris.org
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org

Reply via email to