Gentoo Archives: gentoo-amd64

From: Duncan <1i5t5.duncan@×××.net>
To: gentoo-amd64@l.g.o
Subject: [gentoo-amd64] Re: x86_64 optimization patches for glibc.
Date: Tue, 26 Jul 2005 16:53:02
Message-Id: pan.2005.07.26.16.49.59.697555@cox.net
In Reply to: Re: [gentoo-amd64] x86_64 optimization patches for glibc. by Luke-Jr
Luke-Jr posted <200507261540.06591.luke-jr@×××××××.org>, excerpted below, 
on Tue, 26 Jul 2005 15:40:05 +0000:

> On Monday 25 July 2005 22:38, Olivier Crete wrote: >> On Mon, 2005-25-07 at 22:24 +0000, Luke-Jr wrote: >> > On Saturday 23 July 2005 18:44, Brian Litzinger wrote: >> > > > On Sat, 2005-07-23 at 16:48 +0200, Simon Strandman wrote: >> > > > > Memory to memory copy rate = 1291.600098 MBytes / sec. Block size = >> > > > > Memory to memory copy rate = 2389.321777 MBytes / sec. Block size = >> > > >> > > Memory to memory copy rate = 1302.701782 MBytes / sec. Block size = >> > > Memory to memory copy rate = 2051.979980 MBytes / sec. Block size = >> > >> > Before: Memory to memory copy rate = 557.960449 MBytes / sec. Block size >> > = After: Memory to memory copy rate = 1120.773804 MBytes / sec. Block >> > size = >> > >> > Anyone have a clue why I'm getting half what everyone else gets? o.O >> >> What kind of cpu/ram/motherboard do you have ? > > RAM: 2875MB/1002MB (286%) used (I didn't see swapping during the test, tho) > Motherboard: Asus K8V-SE > CPU: AMD Athlon(tm) 64 Processor 3200+ (2202.876 MHz)
Perhaps it has something to do with the layout of your memory (and BIOS configuration), single-channel vs. dual-channel memory access? I'm getting the same lower (550-ish pre- haven't tried the patch yet) readings here. However, I know I only have a single 512 meg stick slotted for each of my two CPUs (dual Opteron), and I'm running NUMA mode, so the memory is only being accessed at single-channel speeds. I expect I'd double those numbers if I had paired sticks operating in per-node interleaved dual-channel mode. If I turned off NUMA and set inter-node interleaving as well, with 4 matched memory sticks, to get full quad-channel 128-bit interleaving, I expect the numbers would rise accordingly. (Of course, the latter would be at the expense of allowing parallel threads running on each CPU independent but parallel access to their own memory. I can get dual channel without foregoing that, since each node is dual-channel capable, but couldn't get quad-channel without foregoing that independent parallel access, since quad-channel is inter-node interleaved.) Of course, that's just supposition, here. If those with the 1100/2200 rates would confirm whether they are running in dual-channel memory mode, as I suspect, and you confirm that you are running single-channel mode, as I am, it'll pretty much confirm that supposition, however. -- Duncan - List replies preferred. No HTML msgs. "Every nonfree program has a lord, a master -- and if you use the program, he is your master." Richard Stallman in http://www.linuxdevcenter.com/pub/a/linux/2004/12/22/rms_interview.html -- gentoo-amd64@g.o mailing list

Replies

Subject Author
Re: [gentoo-amd64] Re: x86_64 optimization patches for glibc. Sami Samhuri <sami@××××××××××.com>
Re: [gentoo-amd64] Re: x86_64 optimization patches for glibc. Matt Randolph <mattr@×××××.com>
Re: [gentoo-amd64] Re: x86_64 optimization patches for glibc. Raffaele BELARDI <raffaele.belardi@××.com>