Date: Tue, 7 May 1996 17:32:33 -0700 From: asami@cs.berkeley.edu (Satoshi Asami) To: wscott@ichips.intel.com Cc: bde@zeta.org.au, current@freebsd.org, nisha@cs.berkeley.edu, marc@bowtie.nl, ken@area238.residence.gatech.edu, wollman@lcs.mit.edu, pattrsn@cs.berkeley.edu, culler@cs.berkeley.edu Subject: Re: more on fast bcopy Message-ID: <199605080032.RAA10428@sunrise.cs.berkeley.edu> In-Reply-To: <199605071845.LAA03457@ichips.intel.com> (message from Wayne Scott on Tue, 07 May 1996 11:48:28 -0700)
next in thread | previous in thread | raw e-mail | index | archive | help
* The problem is that I gave you numbers on a lower performance desktop * machine. * * On a high performance box with 4-way interleaved memory I get the * following: * * size bandwidth * 32 46.376812 MB/s : * 8192 71.111111 MB/s I see. Well, I'm sorry to be so dense but this is still less than the 80MB/s we are seeing on our P5-133 + Triton boxes. * When I first ran then benchmark the results were toggling between * 66 MB/s and 80 MB/s. The problem was that the benchmark is too fast for * gettimeofday to return accurate numbers. Increasing the size of the * buffer to be copied from 4 MBs to 32 MBs helpped alot. (For those * short on memory just put a loop around the calls to bcopy to make the * time longer.) I suspect this problem was occuring with luke's Pentium * number. There was a problem in the previous version that you tried, the empty() loop we stuck in was subtracting away too much of the time. I disabled that part (so we are now seeing the real speed including the function call overheads) on the version you can pick up from http://stampede.cs.berkeley.edu/~asami/Td/bcopy.html so please try it. Now we consistently get 80MB/s on luke (-current) and two other Pentium 133's (running -stable and -release). (4MB) / (80MB/s) is about 1/20 s or 50 ms, and gettimeofday() returns numbers in microseconds so the granularity shouldn't be a problem. By the way, I made a new version of the scripts, it will now automatically generate .gif files for you if you have gnuplot and ppmtogif. Please grab it (the link is in the web page) if you want to see how I generated the million graphs. :) Satoshi
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199605080032.RAA10428>