On Wed, Aug 17, 2005 at 07:04:52PM -0700, Dan Price wrote: > > [resend, fixing attachment] > > Ok. Attached is a preliminary DEBUG vs. non-DEBUG comparison; I > eliminated the one test which was very obviously misfiring. It could be > that there are others which are also misbehaving. > > For those not acquainted with what I'm measuring: there are two ways to > build the Solaris kernel: DEBUG (i.e. with various debugging checks > enabled, ASSERTs, etc), or non-DEBUG (i.e. go fast). > > We're often asked "What is the overhead of running DEBUG?" to which the > answer is always "it depends on the workload"... and it does. But I > thought it would be neat to get some measurement of how and why running > DEBUG might impact performance. > > As a reminder: we consider benchmarking applications or perf. fixes > on DEBUG bits to be invalid in general. > > If people care, I could post this comparo to the website as well.
This is very cool; one additional point of interest might be "non-DEBUG w/ kmem_flags = f", so that the kmem debugging overhead can be factored out. Cheers, - jonathan -- Jonathan Adams, Solaris Kernel Development _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
