CPU versus Sieving Speed

I would like to know how processor type (and number), memory size, FSB speed and Operating System affect the sieve client speed. I know this is the age old question of allocation, but some of us are new to sieving and there are a few new processors out there now. Anyone?

P2-400Mhz 256Mb 66? FSB Win98se 105kp/s
P3-550Mhz 256Mb 100 FSB Win98se 130kp/s
P3-550Mhz 512Mb 100 FSB Win98se 140kp/s
P3 Xeon
P4
P4 Xeon
Athlon MP
Athlon XP
Athlon 64
Athlon FX 55+
etc.

Stromkarl

BTW, these are using the 1M-20M datfile.

It is important to mention the dat file used, as that affects the speed.
With the 991-50M dat:

Sempron3100 @ 1800 Mhz 200 FSB 512MB - 480kp/s
Athlon64 3400 @ 2400MHz 200 FSB 1024MB - 640kp/s
Celeron300 @450 MHz 100 FSB 256 MB - 109kp/s
AthlonM2400 @1800MHZ 166?FSB 512MB 287kp/s

Athlon MP 2600+ @ 2.0GHz 133FSB 1024MB - ~540kp/s (when I leave it alone, 2 instances running)

P3 1.2GHz(Tualatin Celeron - 256K) 100FSB 640MB - 325kp/s
P3 has async memory PC133 -> 100FSB.

Both these machines use the 1M-20M data file.

Athlon XP 3000+ (Barton) @2.16 (166x13), 512 MB -640kp/s
Using normal and recent 1M-20M dat

Ok... Looks like the AMD guys have it!! :cheers:

Maybe I should have asked the experts how each of these items affects the client?
Should we also be considering the L2/L3 cache? The OS? I will fill in the details I know about if everyone else will.

Stromkarl

Athlon 64 3000+ (939) 1.8 GHz, Linux 2.6, 256 MB, CMOV: 520 kp/s @ 729T
Athlon 64 3000+ (939) 1.8 GHz, Linux 2.6, 256 MB, SSE2: 560 kp/s @ 729T

Question: Because I have SSE2, but it's an Athlon 64, should I do sieve or p-1?


For Reference (Although, the data is from much lower T ranges...):

http://www.free-dc.org/forum/showthr...&threadid=6795

Royanne,

I would personally recommend sieving with the 991<n<50M dat file, using sse2, on the A64, everything other than P4's should be sieving at the moment.

hehe, just realized that I reposted that information yesterday or the day before... thanks for the reply.

I did some extra experiments on the previously mentioned Barton 3000+
Here are the results:
System running Win XP Pro SP1, 512MB 333 DDR (single chip), nForce 2 chipset, CMOV version, running idle immediately after reboot but with all standard programs running in the background (antivirus, firewall etc), 1-20m Dat @ 775T.
Machine is completely stable.

Got 20 consecutive speed reads and averaged them.

Stock system: 2162,72 MHz (166.3 X 13) -> ranging from 636 to 656 kp/s, averaging 646 kp/s

Then I tried a minor overclock using the motherboard utility from within Windows (it is against my wishes to overclock the machine, but just for the sake of experimentation)

2210,37 MHz (170 X 13) -> ranging from 654 to 671 kp/s, averaging 662 kp/s

In this case, the results are proportionate. Overclocking by a ratio of 1,022 gave an increase of 1,0247 in performance.
It's a pity really that I don't have 2x256MB chips, to test performance with half the memory or with dual channel enabled. I'd really like to know how these affect the speed :|
What I noted is that proth sieve is really sensitive to anything other running actively. Even having just the motherboard utility open decreased the speed to about 625-630 kp/s and leaving an idle browser open steals almost 100 kp/s, so anyone else benchmarking their machines should really follow similar procedures to get any results close to realistic (ie, if you want the best your CPU has to offer, let the PC run truly idle with no open program windows)

I'd like to hear other people's views on this and if you think I missed something, I'd love some feedback! Hope this helped...

Maddog1,

Your results seem to vary by quite a bit.

I think what you have to do is let the client run for a while then average 10 results after it has settled down.

Basically look at sobstatus.dat, what happened over night etc.

Comments:

The 100 kp/s decrease in sieve speed from opening a browser window sounds a little odd, perhaps for one itteration this is reasonable but a constant 100 kp/s decrease means that something weird is going on. You may acutally have some heavy spyware etc.. Try running adaware or spybot...

On memory...

I basically only saw a 5-7% increase going from 1 stick of 256 to 1Gb of memory, 2x512MB in my nforce (200fsb).

I was thinking I might actually try some benchmarks using asyc memory settings in the bios.

The other thing you could do is use the two sticks in banks 2+3 for single channel operation, as opposed to 1+2 or 1+3 which would give dual channel.

Just some thoughts check out the new benchmark page.

VJS,
I think the idle results are pretty constant. I did exactly what you described, but averaged 20 results instead of 10 to get the speeds I posted above. I am confident these are accurate. 646 avg on stock speed (636 min-656 max) and 662 avg on 170 FSB (654 min-671 max)

There's definitely no spyware or anything malicious on this machine, I check it constantly (being a bit anal about security!)

Yet, I think I found the guilty part for the strange loss of speed :)
I left the browser idle, but it was on the forum's reply page, which has a BILLION animated gifs for the smilies! These are definitely guilty, as leaving the browser idle on any other simple html webpage, gives back a much lower loss of speed (goes down to 600-630 kp/s values, never lower)
I've been able to witness this even right now, while I'm typing this answer! Speed is much slower than what it was while I was simply looking at the forum threads! I have also seen a similar effect on pages with Shockwave Flash objects, but it didn't occur to me that animated gifs could do the same...
Well, seems I either have to only post rarely on the forum, or learn to type really F-A-S-T :D

Or you have to disable images. :D

With Opera, it's just the "g" key (resp. SHIFT+i since Opera8). Comes handy if you're on 56k dial up...

GIFs don't kill my speed too much, it's the Flash animations... as for the varying sieve rates, it really depends on the ranges it seems. Certain blocks of sieve have some crazy numbers sometimes (on stable systems), probably due to places with few unique factors.

If that doesn't make sense, here's my guess at explanation: You have 10M per iteration, now in those 10M, the sieve only tries the numbers in that range which are prime as any composite numbers would have been tested previously at lower ranges of P. Some ranges may have fewer primes (it's probable rare, but common enough to notice). This would give a shorter factoring time for that iteration, but still the same range of P. Is this not correct?