------- Forwarded Message From: Jim Mauro <[EMAIL PROTECTED]> Subject: Re: (fwd) Re: [perf-discuss] Project proposal: CPUfs To: Richard Elling <[EMAIL PROTECTED]>
I respectfully disagree (I think :^). I think we can improve the measurement, and more compelling, I think our customers should be able to ask the question "how much of my available CPU resources am I consuming?", and allow them to get a reasonably accurate answer. IBM did a pass on addressing this issue: http://www.research.ibm.com/journal/rd/494/mackerras.html For the record (and mostly to keep my good friend Mr. Sneed off my heels! ;^), I'm not advocating the use of CPU utilization metrics as the end-all of performance analysis, I'm simply stating it's important to know how busy the CPUs are. In the age of threads/strands->cores->execution units, it becomes a much harder problem, but it's a problem that, at some level, needs to be solved. Thanks, /jim Richard Elling wrote: > Jim Mauro wrote: >> Hey Walter - by all means continue to challenge and question...it >> helps foster >> discussion and get us aligned in our thinking... >> >> These are all good questions, and the onus is on us to come to >> agreement on the >> answers, which will vary (I think) based on the implementation >> details of a >> particular processor. > > I'm with Walter on this one. Looking at the bottom will make less and > less > sense going forward. > > <geezer> > So we had this problem back in 1987 where we built the MPP system but > couldn't > predict performance except for trivial workloads because the distance > from a > compute element to any sort of large memory or I/O. To some degree it > was > deterministic, but faults or bottlnecks in the (asynchronous) fabric > meant > you never really knew how it would react. We kinda dodged the problem by > implementing codes that worked on a wave-front moving through the mesh > -- not > a generic solution by any means. > </geezer> > > The problem remains unknown, so perhaps we need to take a different > approach, > maybe looking from the top down -- which is difficult without changing > source, > but dtrace opens up interesting opportunities. > -- richard ------- End of Forwarded Message _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
