Author: Matt Taylor
Date: 15:39:21 03/20/03
Go up one level in this thread
On March 20, 2003 at 13:14:38, Robert Hyatt wrote: >On March 20, 2003 at 12:11:11, Matt Taylor wrote: > >>On March 20, 2003 at 10:45:33, Robert Hyatt wrote: >> >>>On March 20, 2003 at 06:58:18, Frederic Louguet wrote: >>> >>>>I have a Dual Xeon 2.8 Ghz here (Dell WS450) with 533 Mhz bus, chipset E7505, >>>>1 GB ram (2 x 512 PC2100 dual-channel non ECC non Registered), Windows XP Pro >>>> >>>>ScienceMark 2.0 beta indicates a maximum memory latency of 88.86 ns and a >>>>2865 MB/s memory bandwith. >>>>I don't run Linux and I don't have LMBench but maybe the ScienceMark results are >>>>similar. >>> >>> >>>What about LM Bench? >>> >>>My dual 2.4's, 2.8's have the E7500 chipset and are reporting 150ns (fastest is >>>145) >>>with 400mhz FSB and registered ECC RAM (DDR). The 3.06's here have the same >>>chipset as yours, with the same 533mhz FSB, but latency is _still_ at 145ns >>>roughly. >>> >>>Perhaps it is the registered / ECC ram that is slowing things down, although I >>>would not >>>use non-ECC memory on any critical machine. >> >>Mission-critical, you mean, and Crafty isn't mission-critical... > >Nope, but crafty isn't the only thing I run. Which explains why it is not on >ICC all the >time, or is sometimes on using different hardware. > > >> >>I've got ECC ram, and I've never experienced a failure in all of the last 9 >>months that I've had it. (Failures generate machine check exceptions which get >>logged for me, and the only machine checks I have logged were generated by >>internal ECC checks on the registers in my Athlon.) > >My cluster has ECC ram. I have had two failures in the past two years. On >Non-ECC >machines, we have had multiple failures that were very hard to detect, because a >program >just crashes but runs fine the next time. Until you stress-test with a memory >diagnostic >that may (or may not) uncover the problem. Yes, I've experienced that. I was boggled one day when the Java compiler took down my machine. A couple months later I got my answer. I also had an interesting experience last April with bad ram. Files in the file cache were being corrupted. Interestingly enough, the OS loaded into the good DIMM because it was lower in memory (since it was in socket 0). The OS was stable, but the applications weren't. I used the same ram in my dual-CPU machine for several months without ever experiencing a failure. >>Non-ECC is faster. I believe it's the registered part that makes it so slow. I >>-think- you can buy ECC unregistered, but most companies won't make it because >>most people want both at the same time. This is what Micron said, anyway. I'm >>pretty sure I've seen ECC unregistered, but it's not terribly popular. There are >>usually restrictions on unregistered ram. I have 4 DIMM sockets and can only use >>2 unless I use registered ram. Most other DDR boards I've looked at have the >>same restriction, from the Abit KG7 to the latest & greatest. >> >>-Matt > >I might investigate this just for fun. I _believe_ my dual has 4 x 256 DIMMS, >but I don't >remember whether it has 8 or 16 slots total (I believe 8 but I'm not going to >open it up at >the moment.) I'll try disabling the ECC on my ram. There isn't anything I can do about it being registered, and I can't tweak any of the settings to get a fair registered vs. non-registered on AMD 760MPX benchmark. This presents an interesting opportunity to see how well my RAID controller works with Linux. -Matt
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.