Welcome Back SpeedRaider!

[tim]

It's good to have somebody else working on this. You've set a good pace.

I see you registered with SOB in in 2006 and got going again this month. What did you do in the interim?

[guru]

I've been busy with OGR-28. I threw a few systems on SOB for testing. I quit a long time ago due to client stability issues. It looks like the client still sucks in that regard. My newer systems crash hard if I try to run the client. Also my output doesn't match the work I do. I added a 12 core systems a few days ago and my output dropped. I have some 8 core systems which the client detects as 4 core with HT. Core 2 never had HT. It would be nice of someone with real programming experience would re-write the client.

[tim]

I wish I had the problem of running 8 or 12 core machines. Quad cores give 0 problems. The last reliability problems I had were long in the past, trying to run it on win95 machines. Yes, that's been awhile ago. I do notice the client runs way faster on intel than amd.
For as long as you want to stay, you're welcome. Folks from free-dc got this team into the top 5 years ago, and since I came relatively late, I've been trying to keep up. Much of what I like about this is the client, that with my machines, it's stable.

[guru]

How about the stats reporting? Is it delayed ore averaged somehow? One thing I like about dnetc is the ability to run a proxy and get stats from it. I can track the performance of each machine to make sure it is working properly.

[pinhodecarlos]

First turn off HT. Then go to local.txt and update:

WorkerThreads=4 (change 4 to how many physical cores you have, do not count hyperthreaded cores)
ThreadsPerTest=1

[tim]

Whoa! with guru on the job, yesterday Free-DC had the top work rate of all teams!

The stats are updated daily. When you log in, you can look at "current pending tests management" or something, and it will show you progress on every test you have pending, from the last time the client contacted the server. One entry for each test. If the date last reported is today or late yesterday, that means your clients are running. That doesn't easily link back to which computer that test is from, though. For me, I have a batch file I run over the network to copy all the prime.log files to one place, then a text processing program grabs the date/time of the last time that machine's client talked to the server, and shows the computer and the date and time. That's over a windoze network. Back when I was doing seti@home classic, I was all linux, and ssh'd in to each, got the same info, and relayed it back to me (all in scripts). It's not as convenient as it could be, but it was fun to write the batch files and stuff.

[guru]

First turn off HT. Then go to local.txt and update:

WorkerThreads=4 (change 4 to how many physical cores you have, do not count hyperthreaded cores)
ThreadsPerTest=1

The client already compensates for HT and only runs on physical cores.

[guru]

Last night I shutdown 4 clients systems(2 are very powerful systems for a total of 30 cores) and today my scores went up. I haven't added or removed any systems in over a week. This non linear reporting of work done is preventing me from throwing a huge amount of CPU cycles at SOB.

Two days ago the stats showed me ~804 T jEMs and a rate of 41 T jEMs/day. Today I am at 1.07 P jEMs. Something is not right with the stats reporting.

Example:
At the top of my stats page it reports Numeric Statistics
Statistic Yesterday
Work rate 321.92 M jEMs/sec

Lower down on the page it reports Neighbors: Last Day's Rate
Username Rate (jEMs/sec)
SpeedRaider 262.384 M

[pinhodecarlos]

I think mprime works differently than the old SOB client. A candidate tested takes at least one week, you will get the credit at the end of the calculations, not intermediate credit as the old client. I could be wrong....

[guru]

My 12 core system completed 5 calculations yesterday but no increase in daily output. If the results were delayed I should have seen an increase.

Overall I on a scale of 1 to 10(highest) I would give the following ratings.
Client stability:7
Client features:9
Documentation:3
Stats reporting:2

[guru]

Over the past 24 hours I shutdown 28 cores (total of 94.5Ghz) with only a 13% drop in daily production output. That left me running 22 cores (62.5Ghz). The problem is a 60% drop in processing power only resulted in a 13% drop in output. The only thing I can think of is a problem with the configuration files. Can someone share the prime.txt file they use?

[tim]

Here's one from a quad core i5, for what that's worth. Don't have any more than quad.

V24OptionsConverted=1
SendAllFactorData=1
StressTester=0
UsePrimenet=1
DialUp=0
V5UserID=spongetim
AskedAboutMemory=1
WorkPreference=0
WGUID_version=2
Windows95Service=1
Left=80
Top=134
Right=1280
Bottom=984
W1=0 0 1182 154 0 -1 -1 -1 -1
W2=0 154 1182 308 0 -1 -1 -1 -1
W3=0 308 1182 462 0 -1 -1 -1 -1
W4=0 462 1182 616 0 -1 -1 -1 -1
W5=0 616 1182 771 0 -1 -1 -1 -1

[PrimeNet]
Debug=1
ProxyHost=
MersenneIP=www.seventeenorbust.com

[guru]

Thanks Tim. It looks almost exactly like mine. Thanks pinhodecarlos for the threads mod. I got it to work by putting it into the local.txt as the prime.txt didn't use it. I do have a question about two entries in the local.txt

12 core 3.47ghz:
RollingCompleteTime=31747118
RollingAverage=857

6 core 3.3Ghz:
RollingCompleteTime=1570060
RollingAverage=944

8 core 3Ghz:
RollingCompleteTime=625436
RollingAverage=850

4 core 2.8Ghz:
RollingCompleteTime=5535789
RollingAverage=2149

Does anyone know if these are related to performance and if so how.

[pinhodecarlos]

Read this thread: http://ftp.mersenneforum.org/showthread.php?t=8652

[guru]

Thanks, that helps explain it. I shutdown all my systems and brought up just the ones I had shutdown two days ago. I found two options that help me run the client on the 8 core (core2) systems properly. Five 8 core systems, one 12 core system and one 6 core system.

[pinhodecarlos]

One thing I forgot, only run intel processors with AVX availability so you can take advantage of this instructions. Please don't run core2 processors.

[guru]

That would be great if the client didn't crash so much on AVX systems.

[pinhodecarlos]

Are your systems stable? Why don't you try to make a torture test with Prime95?

(dnet is not as demanding as Prime95)

[guru]

Yes, my old i7 2600K has been stable for the past 4 years, running everything I can throw at it. Including dnetc for months on end. I'm a software and hardware engineer. I stress test and repair systems for a living.

[pinhodecarlos]

How do you stress test your machines?

[guru]

Many different ways, some of the software is done in house and confidential. Beyond just normal stressing we test the equipment while operating at different thermal limits.

[guru]

The system normally has the turbo set to 45 and is stable up to 50. I tried stock speeds as well as lower than stock. I suspect it is not detecting the processor properly just like the Core2. Core2 has never had HT yet the client assumes HT if more than 4 threads. The 2600K supports AVX but it may be trying to use more advanced instructions like AVX2. Windows blue screens within 30 seconds no matter what speed the system is set to run. I could do further testing by forcing instructions via client options but the system is currently unplugged. Issues like this is why I give the client reliability such a low score. Dealing with issues like this on one system is ok but try running this on 10 or more systems. I don't have that much spare time. I can run something more stable like dnetc and have time to relax, smoke a cigar, and sip some nice Irish whiskey.

[pinhodecarlos]

Do you like BOINC? Care to try http://escatter11.fullerton.edu/nfs/ ?
HT on on this project. Set

lasieve5f - used for huge factorizations, uses up to 1 GB memory: yes

at
http://escatter11.fullerton.edu/nfs/...subset=project

You need at least 1.3GB per thread.

Forgot to add, set this

Swap space: use at most 90% of total
Memory: when computer is in use, use at most 90% of total
Memory: when computer is not in use, use at most 90% of total

at http://escatter11.fullerton.edu/nfs/...?subset=global.

[guru]

For me I do like some control over the clients. I like command line settings and execution, must have proxy support, and reliable stats for correlating clients to work output.

[guru]

Ok here is my real frustration with the SOB stats. Yesterday I shutdown most of my clients and the rate for the day was ~1.3 M jEMs/sec in the 7 day graph. Just a few hours ago I fired up the rest of my systems I was running before. Now the rate for yesterday jumped to ~ 2.6 M jEMs/sec. Stats need to be fixed with respect to time. At a specified time and date the amount of work done needs to be static. Only then can you track your performance.

[guru]

I think I've figured out how the graphs work now. I'm not fond of it but I at least can make some sense of it. Time to turn up the heat. 24 AVX cores and 4 AVX2 cores added.

[tim]

Great!

If you were building a box today, which avx2 cpu would you use? Whatever a guy can afford?

[guru]

It depends on the purpose. Is there a specific application you are running(other than SOB) that needs avx2?

[tim]

Just SOB, the Prime95 client. I sometimes use my boxes for LAN gaming, but seldom. I was curious because I think it would be fun to look at building again. I would guess that going forward more projects would take advantage of the newer instruction sets.

[guru]

Well in that case I think it mostly comes down to cost. However with some instructions the overall processor speed has less of an affect on performance. I'll do some SOB benchmarks on my systems and get back to you. I'll also try dropping the speed on the AVX2 system to see how much it is affected vs AVX only systems.

[tim]

I'd appreciate that, you know more about cpu architecture than I do. *No hurry.* I've seen that intel kicks amd in sob and seti classic, so it would be good to have the latest on what cpu makes most sense.

[guru]

I found PrimeNet has a huge database on CPU's and the client benchmark. The problem is too many misconfiguration systems in the world that will misreport results. I'm going to test a few different architectures and speed then put the results into a spreadsheet. I will try to finish it tonight. I'm curious myself what the results will look like.

I did a quick search through the database and I would say best bang for the buck is the i5 2500k. It has AVX and the third generation i5 wasn't any faster. The 4th generation has avx2 and FM3 which is faster by maybe 25% but my guess is the cost is ~100% more. Provided you can still find the older 2500k. I'm sure they are allover Craigslist.

[tim]

Luckily, I think, I just added 2 i5 2400s. It'll be good to see what you find.

[pinhodecarlos]

The best option for Prime95 or LLR is using an Intel Core i5 4690K with DDR3 running higher than 2400MHz.

[guru]

The best option for Prime95 or LLR is using an Intel Core i5 4690K with DDR3 running higher than 2400MHz.

Based on what? Single socket system? I'll put a dual socket E5 2697 V2 against it. Only 2.7Ghz but 24 cores.

For the cost of the i5 4690K you can build two 2500k systems and get more work done.

[pinhodecarlos]

http://www.primegrid.com/forum_thread.php?id=6126

http://www.primegrid.com/forum_thread.php?id=5524

[tim]

Thanks for the help and pointing out where I can learn more. I'll investigate the i5 4690k. I'm lazy, maybe refurb units will start showing up soon.

[guru]

http://www.primegrid.com/forum_thread.php?id=6126

http://www.primegrid.com/forum_thread.php?id=5524

"So how do they compare? The Sandy Bridge box would do about 414 PPSE units a day if I don't disturb it. The Haswell box would do 666 units a day, about 61% more!!!"
"Some Mega data overnight... the Haswell box is doing about 29% more of them in a given time so a smaller gap than for PPSE (where it was 61%). CPU clock normalised, my Haswell is 41% faster on Mega (76% for PPSE)."

So two 2500K would be 100% faster than one. Given they can be found much cheaper two is better than one.

[pinhodecarlos]

http://cpuboss.com/cpus/Intel-Core-i...-Core-i5-2500K

You make a mistake, you are only considering the investment, you also need to consider the production per watt consumed. Core-i5-4690K is the king for prime search (LLR).

[guru]

Here is some of the benchmark data. Trying to format it in these forums is a PITA! I had to remove parts of the 12 and 24 core benchmarks so it only includes the data for max threads which is all that is really important. The Haswell is really fast at smaller FFT's but once the size starts increasing it's advantage shrinks. This is where the dual socket server CPU's with large cache and more memory channels starts to shine.

i5-4690K
6 MB cache
3.9Ghz x 4 core
25.6 GB/s memory bandwidth
i7-4770K
8MB cache
3.9 Ghz
25.6GB/s memory bandwidth
CPU speed:	3255
Cores:	4
FFT	1024K	1280K	1536K	1792K	2048K	2560K	3072K	3584K	4096K	5120K	6144K	7168K	8192K
1 Core	4.1	5.436	6.788	7.978	9.13	11.65	14.085	16.743	19.169	24.32	29.306	34.888	40.191
2 Cores	2.254	3.06	3.584	4.297	4.975	6.385	7.651	9.158	10.587	13.358	15.888	19.095	21.833
3 Cores	1.586	2.206	2.728	3.325	3.97	5.076	6.125	7.349	8.495	10.694	12.846	15.381	17.458
4 Cores	1.234	1.849	2.479	3.125	3.75	4.807	5.859	6.911	8.041	10.098	12.143	14.387	16.367
i7-2700K
8MB cache
3.9 Ghz
21 GB/s memory bandwidth
CPU speed:	3492.1
Cores:	4
FFT	1024K	1280K	1536K	1792K	2048K	2560K	3072K	3584K	4096K	5120K	6144K	7168K	8192K
1 Core	5.759	7.37	9.022	10.84	12.175	15.597	19.038	22.661	26.39	33.977	41.178	48.555	56.173
2 Cores	6.544	8.429	9.554	11.372	13.686	17.32	22.508	27.162	27.815	35.12	42.799	50.112	58.786
3 Cores	3.028	3.968	4.82	5.834	6.707	8.566	10.326	12.888	14.086	17.801	21.273	25.715	29.768
4 Cores	2.115	2.877	3.564	4.37	5.132	6.692	7.913	9.573	10.962	13.672	16.459	19.577	22.258
Xeon X5690 x two sockets
12 MB cache each / 24MB cache total
3.73 ghz x 12
32 GB/s memory bandwidth
CPU speed:	3466.63
Cores:	12
FFT	1024K	1280K	1536K	1792K	2048K	2560K	3072K	3584K	4096K	5120K	6144K	7168K	8192K
12 Cores	2.819	2.639	3.078	3.991	4.255	5.234	6.582	8.883	8.711	10.789	13.658	21.465	20.025
E5-2697V2 x 2 sockets
30 MB cache each / 60MB cache total
3.5Ghz x 24 cores
59.7 GB/s memory bandwidth.
CPU speed:	3055.28
Cores:	24
FFT	1024K	1280K	1536K	1792K	2048K	2560K	3072K	3584K	4096K	5120K	6144K	7168K	8192K
24 Core	1.306	1.645	1.705	1.922	2.166	2.652	3.011	3.637	4.099	5.526	8.395	12.65	16.836