What machines to run on
14 Nov 2007 0:18:26 UTC
Topic 193313
(moderation:
I am running Einstein on two PC's PIII, 1gig CPU, 512 DRAM. I had no idea the tasks were so long. I have other projects running also on these two machines.
The machines runabout 30 hours a week.
Is there any sense to run Einstein under these conditions?
What machines to run on
)
Assuming that the machines complete their WUs before their respective deadlines and you don't have any computation problems with the WUs in the meantime, they will receive their allotted credit upon completion.
I personally like to see progress more often than that which sometimes drives me to other projects such as SHA-1 Collision Search.
But ultimately it comes down to whether or not you are completing WUs before they are overdue.
Drew A. Withers
http://drewwithers.com
RE: I am running Einstein
)
Welcome to Einstein@Home!
Well, at least the PIII supports the SSE instruction set, so you can use the specially optimized apps for SSE. See this article.
There's still some optimization potential, so if you want to stay with Einstein@Home, you will see the performance of the app improve over time. I think it's not unrealistic to expect that a typical WU will one day require less than 22 h on your hardware, so a WU would finish within a week, well within the deadline of 18 days.
CU
Bikeman
RE: I am running Einstein
)
Like you I tried on an old P3 @ 930Mhz, after seeing how long it was going to take I detached the P3 from the project. I now only use P4's with the Power Users app for einstein.
BOINC blog
RE: RE: I am running
)
Hello!
Just for the records, all the power apps published so far (as well as the most recent Linux stock app) for the intel platform will be able to use SSE on PIIIs. If you are concerned about your electricity bill, you may find that fast PIIIs (the last one built had a 1.4GHz clock speed) will outperform many P4 systems in terms of credits per kWh. For example I have Dual systems that get about 700-800 credits per day on 130 W, which is not so bad for a 6 year old box.
CU
Bikeman
To be honest here, these days
)
To be honest here, these days it really isn't worth running Einstein on a PIII. I have switched 3 of my PIII's off and only run P4's now. Of course if money is not a factor then by all means run PIII's, any RAC at all will help the project.
But with new dual core machines now available for such reasonable outlay it just doesn't make economic sense to run the older kit anymore. At least not on Einsten anyway, it may pay off on Seti.
Waiting for Godot & salvation :-)
Why do doctors have to practice?
You'd think they'd have got it right by now
RE: To be honest here,
)
Being a PIII fan, I can't disagree more :-). In fact, your P4 (no hyperthreading) currently takes more than 120k sec per WU, my PIII only about 60-70k on average. Why? Well, to be fair, That PIII is running the Linux stock app which already supports SSE, while your P4 runs the Windows stock app without SSE optimization. I guess with the current Windows power user app, your P4 still would not be able to beat a .
CU
Bikeman
RE: I guess with the
)
That's absolutely correct. A P4 needs to be running at 2.0GHz at least, to be able to beat a PIII at 1.4GHz. When you also take into account the fact that Tualatin PIIIs run cooler and use less power it can be quite a retrograde step to shut down a Tualatin PIII and replace it with a low to mid-range Williamette or Northwood P4.
And you don't need server grade chips with their larger L2 caches to get the sort of performance you are seeing. This results list belongs to a Celeron 1300 that overclocks quite nicely. It's running at around 1600MHz and is quite stable. As you can see, it produces crunch times in the 50-70K range. The whole machine cost me $10 at auction.
Cheers,
Gary.
RE: But with new dual core
)
Certainly true if you are able to pay up front for the asset in order to save down the track on running costs per unit of work returned. However, what about the situation where you are working to a budget and you already own a decent machine for your normal work. You enjoy seeing the contribution your main machine is making and the opportunity to improve your contribution by making use of somebody else's castoff comes along. You can't justify another full blown purchase of a new machine but you realise that you could easily pay for the electricity used by a castoff by making some minor adjustments to your existing budget. The running costs of a Tualatin PIII desktop are quite modest after all. The machine I linked to in my previous post has a 170W PSU and it's probably only using about half that amount for crunching. I reckon it costs me about $0.50 per day to run it for a RAC of well over 300 - certainly not to be sneezed at :-).
This I don't understand at all. If it "pays off" at Seti, it must also "pay off" at Einstein. Take the Celeron 1300 from my previous post. It completes an E@H task every 17-18 hours or so and gets paid 320 credits per day. If I put it on Seti work, no doubt it would complete a lot more tasks in a day at a lot less credit per task. At the end of a day on each project, it should be paid around the same total amount - actually a bit less from Seti because it's running the "power user" app here. Even if I went for the best current optimised app over at Seti, it probably still wouldn't be quite as productive over there at the moment. If you compared stock app with stock app at both projects, it would probably return about the same - as it should.
So why do you think there might be more justification for running my "old clunker" over at Seti rather than here? :-).
Cheers,
Gary.
RE: RE: But with new dual
)
Gary - you obviously have not been following the thread discussing porting the Alex Khan optimised Power-Mac code to windows.
ATM the most optimised SETI client code is that of Crunch3r using Lunatics V2.4 SSSE3 stuff. The AK code port is showing promise of running SETI at an average of 60% faster than Crunch3r's optimised code. This means a conservative 50% increase in out put on the new Intel Core 2 rigs is possible.
This does not negate the discussion on running a P3 level rig for E@H. I use the SSE optimised 4.36 E@H client on this dual P3 to get an RAC of 450-475. This is backed by a dual Prestonia Xeon, with HT, using the same client. The latter gives an RAC of near 1100.
So I am betwixt and between in this argument, with a lean towards retiring the P3s.
Shih-Tzu are clever, cuddly, playful and rule!! Jack Russell are feisty!
RE: Gary - you obviously
)
Hi John - actually I have, and with quite a bit of interest too :-).
Sure, but this is not really relevant to this discussion yet since there isn't anything currently available and no firm timetable either. An app for modern machines might be just around the corner but I suspect it might be a little while until older machines like ours are supported. Is there actually any indication that SSE hardware will be able to be supported? Even if it can be supported, will it enjoy the same level of speedup? And by that time it's also possible that more of Akos' magic may have been incorporated into the E@H app as well.
I was really trying to focus attention on what the majority of participants end up using - the stock app. Most people don't have the time or energy or level of technical expertese to even fully understand, let alone adopt, the best option for their particular flavour of hardware. I was simply making the point that if you could justify using older hardware for one project, you could equally justify it for the other project. I was also making the point that on what is currently available, an older machine has a higher rate of return here. I was not ruling out the possibility that this might change in the future.
I'm not sure I understand this paragraph. Do you regard those figures as good, bad, or indifferent? :). Surprisingly enough, I have very similar machines to those two - the main difference is that the dual HT Xeon is only 2.4GHz rather than 2.8GHz (same family/model/stepping) and the dual Coppermine is 1.0GHz rather than 933MHz. My RACs are quite similar to yours and I'm quite happy with them.
I also have some dual 866 Coppermines which are shared 50/50 E@H/Seti and I've just had a look at the RACs averaged over a group of 5 identical machines. The machines are running the best currently available optimised app at each project. The E@H RAC is 180 and the Seti RAC is 104. These machines were all turned off for a couple of days around two weeks ago and probably haven't quite recovered yet but the relative values should be OK to compare. It would need a very big improvement in the Seti app to just level with the E@H app. I see a similar pattern on a Northwood SSE2 P4 which is also shared 50/50 between the two projects. The gap is not as large but E@H is still a very clear winner.
I assume you are making this decision because you wish to limit your power bill? If that's the case, you would shut down the dual PIII, because it's a Coppermine and they are quite a bit less productive in terms of credit/kWh. The dual Xeon would use quite a lot more power but not twice as much I would guess.
If your PIIIs were faster Tualatins, it might be more of a contest. The Xeon would certainly win in terms of pure RAC and might also in terms of credit/kWh but it would be a lot closer and would need proper power measurements to prove it one way or another. Dual Tualatins can achieve around 750 RAC which is not that far short of what the 4 virtual CPUs of your Xeon are getting.
Cheers,
Gary.