TR Forums

leor wrote:

guest wrote:

How many PPD would between 1.8 and 2 THz of core 2 Xeons generate using the SMP client? Some people have tried to make SMP performance tables but they don't have Xeon info in it.

depends, netburst xeons or core based xeons?

either way it's gonne net you a lot but core based xeons will be about double

All are core2 as stated.

Finally seem to have gotten the dual-core Linux system off the ground; I'm posting this from it now.

I hate flaky hardware... at first I thought I had a bad PSU, then a bad motherboard, then a bad CPU... finally think I've narrowed it down to two flaky DIMMs in my collection of spare parts. Since I thought I had already ruled out bad RAM (by swapping DIMMs), I spent way too much time trying to chase down what I thought were PSU/mobo/CPU issues.

Anyhow, it's on the air now, and the SMP client is running. PPD seems to be about the same as my WinXP SMP box; I was hoping for more, given what people have said about the performance of the Linux SMP client. This system does have a slight clock speed disadvantage though, so I suppose that's still consistent with what others have reported.

As an aside, the onboard video on this mobo (Tyan Tomcat K8E) really blows chunks. Slow (even in 2D), and the image is noticeably fuzzy. Guess I'll go and pick up a middle-of-the-road PCIe video card, given that I also want to use this system to try migrating most of my everyday computing to Linux.

Congratulations on the new box's birth.

And here I thought I'd catch you in 14.7 years (that's what it was saying yesterday at one point, anyway)...

I'm a little confused, though. I thought you were going to make this your main WinXP box (assuming as I am that this is the Opteron 165), though maybe I've forgotten your saying something about moving to Linux as your main desktop. I thought you had to keep with Windows for work like many of us. (I suppose you could VMware it, though...) If you've only started the new box in recent hours, you can't really say what it's output will be unless you're getting the best WU's available, which seem to be the 1760 point ones. They seem to have gone from almost every WU I ever got to almost rare. I went through a week with none, and the 1523 and 1385 pointers take longer than the 1760's to complete, and obviously that means much lower PPD. I just got 2 1760's in a row, though, so that helps.

If you are running the 165, at 1.8 GHz, then you should be under 900 PPD max. on that machine if my calculations relative to my overclock and its output are accurate. I'm running Linux under VMware and you don't have that to put up with, but my calculation was 1800/2561 * 1176 (my approximate PPD on 1760's), and that yields ~826, and I'm fairly confident I'm not losing more than a smidgeon from virtualization. Your output's still good for one machine, no question. But the heat made by the chipset and CPU on my NF4 Ultra board make me want to knock it back down to stock speed (and push it into another room) and replace it with a C2D, but I think I may be spending a lot of money on something else in short order, so oh well.

BTW I haven't been able to get my K7D motherboard to POST, and I can't tell what the problem is. I have new RAM, and the only 2 things I'm not so sure about are the MP 2000's, and the PSU I'm using. I even went so far, though, as using one PSU for the 20-pin connector, and another for the 4-pin 12 volt one, and that didn't make any difference. Next step is to pull out the CPU's and try my known-working Barton in one socket, but I'm doubtful, and I really hate the work involved with the heatsink and Arctic Silver just to prove something I'm fairly confident about. It will let me test both MP's in the other board, I suppose, though. Hmph.

Ragnar Dan wrote:

I'm a little confused, though. I thought you were going to make this your main WinXP box (assuming as I am that this is the Opteron 165), though maybe I've forgotten your saying something about moving to Linux as your main desktop. I thought you had to keep with Windows for work like many of us. (I suppose you could VMware it, though...)

Heh... yeah, it's complicated (and messy). I am still keeping the WinXP box as well, since there are some thing I do need Windows for. This new system was originally supposed to be the Opty 165, but at some point in the troubleshooting process, I became absolutely convinced that the CPU was at fault, and went and picked up a cheap X2 3800+. I suppose I could swap the Opty 165 back in now, but for the time being the X2 3800+ is in there. For folding, the slightly higher clock speed probably doesn't make up for the smaller cache... oh well.

The good news is that it should be feasible to get the Opty 165 going in addition to this one. I have another S939 motherboard, and can probably scrape together enough working DDR DIMMs to make a go of it without shelling out any more dough.

If you've only started the new box in recent hours, you can't really say what it's output will be unless you're getting the best WU's available, which seem to be the 1760 point ones.

True... I'm only basing my estimate based on the first WU it got. I'll have to watch and see how it does over the next couple of weeks.

D'oh I hate discovering my mistakes when they're quoted by others ("it's" instead of the proper "its").

I don't recall if anything definitive about cache has been reported. If so, I've definitely forgotten about it. It must help on certain memory-intensive units, but that's only assumption.

A cheap 939 X2 3800? I checked Pricewatch and they list at just over $100. It's much cheaper than it used to be, but over $24 more than an AM2. Of course that place isn't what it used to be for lowest prices.

If you can get the Opty going, you definitely should if you can justify the power expense. I'd replace something more power-hungry with it if need be. I'm using almost 15 amps in my bedroom now when the room's window A/C is on. I'd love to be able to not run that thing, but I'd have to have much cooler machines, and put them elsewhere. I can probably move the Barton somewhere... if I can find drivers for the wireless USB doohickey under Win2k.

Ragnar Dan wrote:

I don't recall if anything definitive about cache has been reported. If so, I've definitely forgotten about it. It must help on certain memory-intensive units, but that's only assumption.

There's been some speculation that lots of cache really benefits some of the large SMP WUs.

A cheap 939 X2 3800? I checked Pricewatch and they list at just over $100. It's much cheaper than it used to be, but over $24 more than an AM2. Of course that place isn't what it used to be for lowest prices.

TigerDirect currently has 'em for $80 (grabbed one at their outlet store on my way home from work the other day). It's an OEM one though, so if that $100 is for retail (with HSF) it's not too out of line.

If you can get the Opty going, you definitely should if you can justify the power expense. I'd replace something more power-hungry with it if need be.

I may end up retiring the dual CPU Barton rig. It's definitely the most power-hungry of the systems I have here right now.

just brew it! wrote:

As an aside, the onboard video on this mobo (Tyan Tomcat K8E) really blows chunks. Slow (even in 2D), and the image is noticeably fuzzy. Guess I'll go and pick up a middle-of-the-road PCIe video card, given that I also want to use this system to try migrating most of my everyday computing to Linux.

you might want to pull down the ATI video drivers for linux. i think it's weird that they're using an ATI video device w/ an nforce4 chipset... but whatever.

eitje wrote:

you might want to pull down the ATI video drivers for linux. i think it's weird that they're using an ATI video device w/ an nforce4 chipset... but whatever.

Well, I can sorta see the logic to it. Since it's nominally a server board, they didn't need a modern GPU, and didn't want an IGP which was going to steal memory bandwidth from the CPU. Dunno if nVidia has anything appropriate still in production; from what I've seen, ATI seems to be better about keeping older tech around for niche applications like this.

I figured the Rage XL was old enough that Linux would have pretty good support built in. Maybe not; even though it identified the chip correctly, performance-wise it feels like it is using a "dumb" framebuffer driver.

Edit: There do not appear to be any Linux Rage XL drivers available for download on AMD/ATI's site.

Not sure about the Linux ATI drivers, but most ATI drivers are backwards compatible. Although Rage XL might be a stretch. My new/used 9600XT did NOT play well with the 7.4 DAAMIT drivers, but is running fine with 6.14 from my other 9600XT's cd. Good Luck.

Best I can tell after some Googling around is that accelerated Rage XL drivers for Linux do exist, but they are not part of the default set of Xorg drivers because they contain a potential security exploit that nobody's gotten around to fixing yet. So if you want 'em you have to build 'em yourself from source.

I don't think current Radeon drivers are backward compatible with the Rage XL, it is based on a much older chipset (Mach64).

Edit: I'll probably just pick up a cheap, passively cooled nVidia card... something like this. If the image quality of the onboard Rage XL was better, maybe I'd be willing to expend the effort to hack around and figure out how to compile the accelerated driver for it. But given that I'd still be pissed off at the IQ anyway, it just isn't worth it...

just brew it! wrote:

Edit: There do not appear to be any Linux Rage XL drivers available for download on AMD/ATI's site.

Don't worry, even if ATI made some drivers available they'd still probably suck.

just brew it! wrote:

Ragnar Dan wrote:

A cheap 939 X2 3800? I checked Pricewatch and they list at just over $100. It's much cheaper than it used to be, but over $24 more than an AM2. Of course that place isn't what it used to be for lowest prices.

TigerDirect currently has 'em for $80 (grabbed one at their outlet store on my way home from work the other day). It's an OEM one though, so if that $100 is for retail (with HSF) it's not too out of line.

Yeah retail boxed. All I ever check because they have manufacturer's warranties, and that's something I'll pay the difference for even when I mean to buy an aftermarket HSF.

just brew it! wrote:

Ragnar Dan wrote:

If you can get the Opty going, you definitely should if you can justify the power expense. I'd replace something more power-hungry with it if need be.

I may end up retiring the dual CPU Barton rig. It's definitely the most power-hungry of the systems I have here right now.

Sounds like a good idea. The addition of SSE2 alone helps on folding. I must say, again, though, that my 165 makes more heat than any system I've ever bought. I've got the P4 from work which probably makes a greater amount, but I don't count that machine (esp. since I don't run it too much), and I'm not absolutely sure about it making more. This thing makes heat that seems like it's not as hot, but there's more of it. Giant heatsink, and the chipset, and maybe the video card (though it's always in 2D mode) may add more heat.

I definitely need to buy that killawatt thing and find out what I'm spending on this machine, and then try to figure out a way to compute how much extra A/C power I'm using because of it compared to a new C2D, and maybe I can justify it that way... especially since I've almost definitely decided against the other large purchase I was thinking I'd make (Honda bike upgrade). Mostly. Probably.

just brew it! wrote:

Edit: There do not appear to be any Linux Rage XL drivers available for download on AMD/ATI's site.

good call - my fault on not reading the FAQ.

The ATI Proprietary Linux driver currently supports Radeon 8500 and later AGP or PCI Express graphics products, as well as FireGL 8700 and later products.

One problem to report: was working on one WU and it said memory allocation failure. Increasing my VM to 768MB let the WU pass through and hopefully a big chunk of points for UGN. So it looks like 512MB is a bit tight? I have turned the RAM allocation in the VM back down to 600 and we will see if it works, now that I have backup+restore capability.

Going back on (my) topic: after some struggle with my networking noobness, finally I have set up something that allows me to run the diskless SMP client with backup and the ability to restore the backup if necessary. Yay! As fellow TR Gerbil Ragnar Dan noted, one could install a full Linux and then run the SMP client. Then the WU files can be transported off and such. However, due to my laziness to learn Linux, I'm not prepared to do that. I want to remain as much Windows as possible. notfred's excellent diskless setup seems like a relatively simple way to get into Linux SMP under a VM for more points. However, there are only 2 ways to do backup+restore: a) download the backup.tar file from the web interface, but you need the ability to un-tar the files on a real Linux setup to resume, or b) use TFTP and it will automagically do backups, and on reboot will pick up right where the backup file left off.

I have picked (b), but there are a few problems:
- I use a consumer grade router and its built-in DHCP server doesn't dish out the proper information for the automatic TFTP to work (it uses a similar setup to PXE boot, needing extra information during the DHCP lease process).
- OK, I lied in the above , I actually don't use the DHCP server on the router anyway. I had my file server+domain controller serve DHCP request. However, that DC is the key to the home LAN and I don't want to mess it up. So I will need a separate DHCP server, possibly pointing to a new subnet.
- I would also like the Folding clients to seamlessly connect to the internet for WUs and stuff as well. So if I want a separate subnet underneath the home LAN, yup, you know it, I basically need a router for the 2nd subnet.

So after some trial and tribulations (part due to my general networking noobness, part due to me struggling with VMware's networking setup), here it is (ASCII art ftw).

Since I am more familiar with Windows, the Router VM ended up using up more resources than it could have been. It's ok, I'll trim it or use a smaller Linux setup eventually. The 2 VMs live on my machine. The Router VM is set up to have a bridged NIC and a private NIC (the tricky part was to tell VMware not to use NAT nor DHCP on that NIC).

Router VM:
- Windows 2000 Server (those who don't want to dabble with Linux and have some spare licenses lying around, this is for you )
- Setup DNS to forward to the subnet above the router, in my case 192.168.0.10. Those with a more capable software router (could be any OS really) will usually take over 192.168.0.1 and be the one to set for forwarding.
- Setup DHCP to dish out IP addresses with the 2nd subnet, and include the necessary info for PXE boot and pointer to the TFTP server (self). See note 1.
- RRAS: in NAT mode to properly route packets in and out of the subnet. Essentially nodes on the 2nd subnet pass through 2 layers of NAT to get to the internet.
- TFTP: Part of the reason why I chose Win2K over Win2K3. I can use the tftpd that is included, without using tftpd32. I created the folder for tftpd to hold files, then I can get to the files by either using Windows file share or drag 'n drop via VMware Tools (only VMware Workstation has this ability). See note 2.

SMP FAH VM:
- This one is easy, and in theory can be replicated very quickly (that's the whole point) if you are building a farm. I mounted the generated ISO from notfred's site, use the .vmx file from the website, fire it up, and voila!
- Every 15 minutes it will save the backup via TFTP. I check the Router VM and the .A and .B files are both there. Nice.
- Remember the problem I mentioned in the beginning? Well, I just restart the VM after the memory allocation failure, the "new instance" picked up the backup file from TFTP, loaded the backup and restarted, from 93%. Yes!

Notes:
1) I followed the instructions here to set up tftpd.exe and the proper extensions that I need for DHCP. Then I realized I did too much. I could have used tftpd32 and get away with not touching Windows DHCP. Well, lesson learned.
2) The reason why I thought I needed MS tftpd instead of the excellent tftpd32 is that I needed something that can also do NAT+DHCP+DNS. So I took the long route and do them separately on Win2K server. However, I have since found out that tftpd32 can do good enough DHCP. As long as it is in its own subnet, I can mess it up to my heart's content anyway. And you don't need to manually add those extra info for DHCP to dish out.
3) Should you decide to use tftpd32, then all you need is RRAS for the NAT, and DNS so the FAH clients can go on to the internet on their own.
4) 30-50megs a backup is quite an amount of data travelling on my subnet. The RRAS applet showed the incoming and outgoing bytes.

Hope this helps or inspires others for their own setup. I think we need something like an Overclockix that is designed to work with notfred's stuff. What I have in mind is a distro that has these:
- DNS
- DHCP
- TFTP
- SMB
- (optional) its own folding client
Which should help people build their own farm on its own subnet without interfering with other real computers in a home LAN environment. I have installed Damn Small Linux and it's really small, but I don't think I need all the client apps, plus I need to add back dhcpd, dnsd (bind?), iptables and tftpd. The problem with a regular Linux distro is that all are not part of the default and needs some hardcore setup process. If we can build some of those graphically driven (or just ask a bunch of questions in text mode?) screens and it then configures all that stuff, it seems like an easy way to build a farm? /dreaming

Or if someone knows of a less convoluted way to deal with one consumer router and tftpd32 without affecting other nodes on the home "main" LAN, I'm all ears too.

Latest addition to the arsenal:


It's the Opty 165 I mentioned a while back... the Opty 165 build turned into an A64 3800+ build; the Opty 165 has a new home on an Asus A8V-VM, along with some freebie used DIMMs a co-worker was giving away recently.

I'm loading Fedora Core 6 on it as I type this, and if all goes well it should be up and folding shortly.

I picked up something similar to this for my F@H Fedora installs.

eitje wrote:

I picked up something similar to this for my F@H Fedora installs.

I've actually got all the Fedora updates and extras on an external USB drive already. I ought to put the install image on there too.

Had a few minor glitches getting that system going but it's good now. Just downloaded its first SMP WU about a half hour ago.

Hm... I think I'm still too concerned about outside magnetic fields to put a hard disk on top of an optical drive. Foolish, no doubt.

Is it me or is there something different looking about that HD? Maybe it's older than the ones I've seen recently and forget, or a brand I don't use which has that shape on the case.

I've got a cruddy 2608 WU crunching now. So much variation in output, so sad.

Ragnar Dan wrote:

Hm... I think I'm still too concerned about outside magnetic fields to put a hard disk on top of an optical drive. Foolish, no doubt.

The coercivity of the magnetic coating on modern hard drive media is so high you could probably set a pretty strong magnet right on top of the drive and not have an issue.

Is it me or is there something different looking about that HD? Maybe it's older than the ones I've seen recently and forget, or a brand I don't use which has that shape on the case.

It's a ~2 year old Hitachi drive. I've been quite happy with Hitachi drives lately. Whatever the issues were with the Deskstar line, IBM/Hitachi seem to have sorted it out. I'm a little surprised they kept using the brand name though, given all the "Deathstar" jokes!

I've got a cruddy 2608 WU crunching now. So much variation in output, so sad.

Still, in terms of average PPD across the overall WU mix, the SMP beta WUs are hard to beat.

Yeah, I know I'm being foolish about HD's. I hope I didn't word that so as to sound as though I didn't mean it about myself.

As for WU's, I know, I'm spoiled by the 1760 pointers. They kick the crap out of anything else so incredibly much that even the great gain from the one I'm running seems too little. Especially since it won't allow me to finish a new WU before the weekly ExtremeOverclocking cutoff at the end of Saturday.

I actually wasn't all that impressed with the first couple of WUs the Linux SMP box got. PPD was decent, but not terrific. Then, this afternoon it downloaded a Project 2605 WU.

O... M... F... G...

826 PPD.

This actually prompted me to get my live status page working again, so I can easily check current production at a glance. (Before tonight the page was in sorry shape -- the WU database it was using was months out of date, and it didn't handle SMP WUs properly.)

And look at the poor 800MHz Athlon... the A64 X2 3800+ SMP client is out-producing it 40:1!

I think I'm gonna be able to "go red" again for the first time since I turned off the diskless farm!

Yeah, the 2605 are not affected by the size of an L2 cache anywhere close to the 2608/2609/2610 series.
X2s with 2x512k cache also do well on the 2605s.
The benchmark machine has a huge amount of L2 cache so other CPUs that have lots of cache (e.g. 6320/6420/6600) PPD is much more consistent.

I think I've been spoiled with 2605's, because I was getting almost nothing but them for 2-3 weeks. Getting two 2608's (1385 points) and a 2609 (1523) consecutively irked me an amount.

According to the Linux SMP FAQ, they use the same old P4 2.8 GHz w/o hyperthreading to benchmark the WU, though that has to be impossible as far as the final deadline date is concerned. Anyway, that P4 only has 512kb of L2 cache.

But anyway, I'll be looking more closely at C2D's when the July price cuts come. It may be smart to buy memory now, though. Or actually a couple of weeks back, since I've only seen one good deal (after rebate) on fast DDR2 lately and it was from TigerDirect, who I don't trust for rebates since they screwed me out of a big one.

Ragnar Dan wrote:

I think I've been spoiled with 2605's, because I was getting almost nothing but them for 2-3 weeks. Getting two 2608's (1385 points) and a 2609 (1523) consecutively irked me an amount.

According to the Linux SMP FAQ, they use the same old P4 2.8 GHz w/o hyperthreading to benchmark the WU, though that has to be impossible as far as the final deadline date is concerned. Anyway, that P4 only has 512kb of L2 cache.

The benchmark machine for SMPs is actually a modern Xeon (C2D based) system (4 cores) with tons of L2 cache that can blow through a 2605/2608/2609 in under a day. C2D with tons of L2 cache (6320/6420/6600 and higher and T7200 and higher) will also do quite well. Those with low cache (especially X2s with 2 x 512K) but even 6300/6400/T5600/T7100 will feel lots of pain for 2608/2609 (can drop the PPD by half in some cases!). Small cache X2s (2x512K) like mine, when running at 2.4GHz usually can't even make the preferred deadline for 2608/2609 WUs! And not making the preferred deadline means that the WU will be reassigned!!
Hopefully (but sadly I don't expect...) in the future, cache loving WUs will not be assigned to CPUs with small caches.

Tarx wrote:

Those with low cache (especially X2s with 2 x 512K) but even 6300/6400/T5600/T7100 will feel lots of pain for 2608/2609 (can drop the PPD by half in some cases!). Small cache X2s (2x512K) like mine, when running at 2.4GHz usually can't even make the preferred deadline for 2608/2609 WUs! And not making the preferred deadline means that the WU will be reassigned!!

My X2 3800+ (2.0GHz) seems to be doing OK on the 2608/2609 WUs. It's the Linux client though, and yes I've noticed that -- while it does make the deadlines OK provided the machine doesn't have a lot of downtime -- production is substantially lower than on some of the other SMP WUs.

I did have to stop running the SMP client on my Athlon MP box (2x Athlon MP @ 2.0GHz), because it was intermittently missing deadlines.

Hopefully (but sadly I don't expect...) in the future, cache loving WUs will not be assigned to CPUs with small caches.

Even if they implement an algorithm to take cache size into account (and yes I agree they probably won't), it would likely be only a preference, not a hard rule. Otherwise they might end up with a huge backlog of large-cache WUs, since there are probably more systems out there with small caches.

just brew it! wrote:

Tarx wrote:

Those with low cache (especially X2s with 2 x 512K) but even 6300/6400/T5600/T7100 will feel lots of pain for 2608/2609 (can drop the PPD by half in some cases!). Small cache X2s (2x512K) like mine, when running at 2.4GHz usually can't even make the preferred deadline for 2608/2609 WUs! And not making the preferred deadline means that the WU will be reassigned!!

My X2 3800+ (2.0GHz) seems to be doing OK on the 2608/2609 WUs. It's the Linux client though, and yes I've noticed that -- while it does make the deadlines OK provided the machine doesn't have a lot of downtime -- production is substantially lower than on some of the other SMP WUs.

Yeah, my system is used quite a bit on other tasks so that doesn't help, and the Linux SMP runs in VMware that slows it up a bit (10% to 20%) versus native. It does make it for the final deadline, but usually not the much more important preferred deadline for 2608/2609.

Tarx wrote:

Yeah, my system is used quite a bit on other tasks so that doesn't help, and the Linux SMP runs in VMware that slows it up a bit (10% to 20%) versus native. It does make it for the final deadline, but usually not the much more important preferred deadline for 2608/2609.

According to the current project summary chart, the preferred and final deadlines are the same for those WUs (4 days)?

just brew it! wrote:

Tarx wrote:

Yeah, my system is used quite a bit on other tasks so that doesn't help, and the Linux SMP runs in VMware that slows it up a bit (10% to 20%) versus native. It does make it for the final deadline, but usually not the much more important preferred deadline for 2608/2609.

According to the current project summary chart, the preferred and final deadlines are the same for those WUs (4 days)?

Wow! Looks like they've matched the preferred & final deadlines for 2608/2609 (but not 2605, but that shouldn't be a problem for an X2)! There was a thread about this a week or so ago that was discussing this problem (about preferred vs final for the Linux client). Off I go to restart the client

Has anyone tried SMP folding on the new Pentium E21x0 series CPUs? The benchmarks on games and applications show them being 20% faster vs. similarly clocked X2s.

I tried searching info on FAHinfo for the Allendale core, but all I get a overclocked scores, nothing at the stock clocks of 1.6 and 1.8 ghz.

http://fahinfo.org/index.php?allscores= ... 0&sort=pna

My gut is that if Presler can meet deadlines, then so can allendale. Anyone know the PPD of these processors? What are your thoughts?

Tarx wrote:

Wow! Looks like they've matched the preferred & final deadlines for 2608/2609 (but not 2605, but that shouldn't be a problem for an X2)! There was a thread about this a week or so ago that was discussing this problem (about preferred vs final for the Linux client). Off I go to restart the client

I still wish they'd bump it to, say, 6 days though. As it is, if the system is used for anything else, or needs to be taken down to install/test hardware upgrades, making the deadline can be iffy.