Date: Mon, 10 Feb 1997 19:54:56 +1000 (EST) From: Stephen McKay <syssgm@devetir.qld.gov.au> To: richard@pegasus.com (Richard Foulk) Cc: freebsd-hardware@freebsd.org, syssgm@devetir.qld.gov.au Subject: Re: 64 MB ECC or 128 MB non ECC ? Message-ID: <199702100954.TAA24699@ogre.devetir.qld.gov.au>
next in thread | raw e-mail | index | archive | help
richard@pegasus.com (Richard Foulk) predicts: >} * From: roberto@keltia.freenix.fr (Ollivier Robert) >} >} * One thing to consider is that you'll suffer a 10-15% speed penalty with ECC >} * RAM. (number from some -hardware mails in the past). >} >} To clarify: 10-15% penalty on maximum memory bandwidth. (E.g., 70MB/s >} vs. 64MB/s on TritonII with 66MHz bus and P5-133.) >} >} This is NOT the same as application speed penalty, which obviously >} varies depending on how memory-intensive it is. > >That's kinda what I thought. So closer to 1%. Indeed, I've just waded into the wake left after the bleeding edge, and bought a Rhino-9 MB with 512kb cache, 32Mb of 60ns FPM parity ram, Hercules Trio64V+ based video card, and borrowed (yes, some people are really gullible!) a Pentium 120. I've been experimenting with the Adaptec 1542B and Wren 6 SCSI disk from my old toy box, and under FreeBSD 2.1 building a kernel with ECC enabled uses about 1% more cpu time than building a kernel with just PARITY enabled. I've done 20 runs, all with fresh reboot, compile, compare, delete, halt. (user+system time is just short of real time, so the slow disk isn't hurting too much.) The variation in individual runs was about 1% and I'm still a little concerned that maybe the ECC BIOS setting isn't doing anything at all. Cynical old me. So, my next set of tests will be with L1 and L2 cache disabled to highlight the speed difference. Much of this difference will be mopped up by the write merge buffers on the TXC, but it should still be bigger than 1%. Final proof that the ECC setting does anything other than slow my machine down will have to wait until I work out how to reprogram the TXC to suddenly switch from PARITY to ECC. This should immediately cause an ECC failure. Stephen.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199702100954.TAA24699>