stats update

junky · 2004-01-23, 02:46

hi, like im new with NSF client, i wanna know some little things.
First, stats are updated once a day ? if yes, on what time ?
is there a way to know my current speed ? i wanna make some tests with friends if there's any possibilities.
is there any webpage talking about benchmark and stuff like that ?

More questions will follow soon :)

TauCeti · 2004-01-23, 09:32

Hello junky,

afaik there is no stand-alone benchmark of the NFS-client available to compare results 1:1. But at least if you use the windows-GUI version of the client, you can compare the speeds easily if some requirements are fullfilled

1) I suggest you let the client run for at least 30 minutes on each test-system. At first the clients only gets small assignments and that leads to lower benchmark values (because of the overhead of reading the roots.txt file)

2) The clients have to work on the same Project. If they also work in the same pool and get roughly similar "start of range" values, the results are comparable.

3) In the GUI there is a value for "lines per day". You can also reset the benchmark (reset button) to start the calculation of that value anew.

I would be most interested in some benchmark numbers here. If we post the project (should be 2_811M for all users now) and the range, we can compare the values.

I think i already have found out some things regarding performance with the windows-client:

a) The client very much depends on memory bandwidth. At least with P4-systems. Contrary to the prime95 client, memory bandwidth is more important then core speed. So NFSNET is a worthy project for dual-channel-DDR owners ;)

b) So far there are no SSE2 optimisations in the client. I guess that the new Opteron / Athlon94 systems really should shine in all glory with NFSNET due to their integreted memory controller and high IPC. But sadly i have no access to such a system.

c) Even older P-III systems work pretty well with NFSNET. My old P-III 933 manages roughly 40% of the speed of my 3GHz Dual-DDR P-4.

d) If you know people who own a HT-enabled P4 and would like to crunch with NFSNET _and_ GIMPS, they should try choosing "trial factoring" in GIMPS and use both clients simultanously. That way for factoring above the 2^64 threshold (SSE2 is used in that range) you can improve the overall performance of your HT-P4-system to about 170% compared to running only one client. But if you do use NFSNET exclusively, you help us even more *g*

Have fun!

Tau

xilman · 2004-01-23, 16:04

Quote:

Originally Posted by junky

hi, like im new with NSF client, i wanna know some little things.
First, stats are updated once a day ? if yes, on what time ?
is there a way to know my current speed ? i wanna make some tests with friends if there's any possibilities.
is there any webpage talking about benchmark and stuff like that ?

More questions will follow soon :)

Stats are normally updated daily, a short time after 01:00 UT. During the winter, that's the same time as GMT, it's 2am Central European Time and 8pm the previous day on the east coast of the US. Add an hour in the summer.

The reason for choosing 1am is that NFSNET works exclusively on UT time. At 00:00 each day the task allocation servers create a new directory and send all incoming information to files held in there. Around 00:15 (a pause to ensure the dust has settled), they then perform various housekeeping tasks, which includes creating a summary of the previous days results and sending it off to the central stats server. The stats server waits until around 01:00 to ensure the summaries have arrived from all the servers before collating them and making overall stats available on the web page.

Sporadic updates to the stats page are made if it's necessary. Occasionally the automated update breaks and this is one reason for us kicking off a run of the stats generation code.

Benchmarking information is sadly lacking at present. We shoud do something about that lack...

Paul

junky · 2004-01-25, 02:42

thanks guys for the info.
im running the client for about 2 days now, and i cant see my stat.
Even if i check at: http://www.nfsnet.org/stats2/statsre...agesize&page=0 , i cant see that im a new user. My user id is: junky.
what's wrong ? do i need to complete a certain amount of work to be register ?

thanks in advance.

Wacky · 2004-01-25, 10:28

In order to track this down, I need quite a bit more information from you.
Please e-mail a copy of the processors/config.txt file to me. I also need to know the pool from which you are getting assignments. Send me a couple of the output lines that say which assignments you have received.

Send all of this to richard@nfsnet.org and I will try to locate your records.

junky · 2004-01-25, 13:57

hi wacky, its now oky.
my stat is at: http://www.nfsnet.org/stats2/statsre...&userid=junky&

thanks for nice support ;)
and i've a question, how much time is required to complete a workunit ?
a workunit is composed of how many lines ? relations ?

and finally, is 3 wu is "fast" for a single day ?

Thanks

Wacky · 2004-01-25, 15:24

Junky asks:
how much time is required to complete a workunit ?
a workunit is composed of how many lines ? relations ?
and finally, is 3 wu is "fast" for a single day ?

Well, first I should point out that in the NFSNET sysem, an assignment is not the same as a workunit. The workunit (WU) was designed to measure the computational effort in a manner that is independent from the particular sieving problem.

In the case of our current project, each line is 0.00113903 WU.
The number of relations generated by each line varies greatly depending on both the project and the particular line. In general the yield (relations per line) decreases with increasing line number. At very low line numbers, we were getting more than 10 relations per line. By the time we finish sieving, the current yield will have dropped to about 2/3 relations per line. Right now, the yield is a little above 1.

As for how long it takes to do a WU, it all depends on your machine. A typical 1GHz PIII processor will produce about 2 WU per day if it runs full time. So is 3 WU/day fast? It certainly is not the fastest that we have seen. I have a single processor machine turning in 4.6 WU per day, and it isn't the fastest machine in the pack. However, you might note that we obtained less than 600 WU from nearly 300 machines yesterday. So 3 WU per day is well above the average. (But a number of the machines put in only 12 hours per day).

junky · 2004-01-25, 16:12

really clear explanation wacky, thanks for the info :)

2004-01-23, 02:46	#1
junky Jan 2004 7×19 Posts	stats update hi, like im new with NSF client, i wanna know some little things. First, stats are updated once a day ? if yes, on what time ? is there a way to know my current speed ? i wanna make some tests with friends if there's any possibilities. is there any webpage talking about benchmark and stuff like that ? More questions will follow soon :)

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
CSVs for stats available + New combined stats	opyrt	Prime Sierpinski Project	3	2010-05-31 08:13
LMH update	monst	Lone Mersenne Hunters	8	2009-01-22 13:33
No update possible	Max	Information & Answers	5	2009-01-20 18:12
update	wfgarnett3	PSearch	0	2005-08-25 09:41
Update	wblipp	ElevenSmooth	5	2004-12-28 18:25

2004-01-23, 09:32	#2
TauCeti Mar 2003 Braunschweig, Germany 2×113 Posts	Hello junky, afaik there is no stand-alone benchmark of the NFS-client available to compare results 1:1. But at least if you use the windows-GUI version of the client, you can compare the speeds easily if some requirements are fullfilled 1) I suggest you let the client run for at least 30 minutes on each test-system. At first the clients only gets small assignments and that leads to lower benchmark values (because of the overhead of reading the roots.txt file) 2) The clients have to work on the same Project. If they also work in the same pool and get roughly similar "start of range" values, the results are comparable. 3) In the GUI there is a value for "lines per day". You can also reset the benchmark (reset button) to start the calculation of that value anew. I would be most interested in some benchmark numbers here. If we post the project (should be 2_811M for all users now) and the range, we can compare the values. I think i already have found out some things regarding performance with the windows-client: a) The client very much depends on memory bandwidth. At least with P4-systems. Contrary to the prime95 client, memory bandwidth is more important then core speed. So NFSNET is a worthy project for dual-channel-DDR owners ;) b) So far there are no SSE2 optimisations in the client. I guess that the new Opteron / Athlon94 systems really should shine in all glory with NFSNET due to their integreted memory controller and high IPC. But sadly i have no access to such a system. c) Even older P-III systems work pretty well with NFSNET. My old P-III 933 manages roughly 40% of the speed of my 3GHz Dual-DDR P-4. d) If you know people who own a HT-enabled P4 and would like to crunch with NFSNET _and_ GIMPS, they should try choosing "trial factoring" in GIMPS and use both clients simultanously. That way for factoring above the 2^64 threshold (SSE2 is used in that range) you can improve the overall performance of your HT-P4-system to about 170% compared to running only one client. But if you do use NFSNET exclusively, you help us even more g Have fun! Tau

2004-01-25, 02:42	#4
junky Jan 2004 10000101₂ Posts	thanks guys for the info. im running the client for about 2 days now, and i cant see my stat. Even if i check at: http://www.nfsnet.org/stats2/statsre...agesize&page=0 , i cant see that im a new user. My user id is: junky. what's wrong ? do i need to complete a certain amount of work to be register ? thanks in advance.

2004-01-25, 10:28	#5
Wacky Jun 2003 The Texas Hill Country 1089₁₀ Posts	In order to track this down, I need quite a bit more information from you. Please e-mail a copy of the processors/config.txt file to me. I also need to know the pool from which you are getting assignments. Send me a couple of the output lines that say which assignments you have received. Send all of this to richard@nfsnet.org and I will try to locate your records.

2004-01-25, 13:57	#6
junky Jan 2004 7×19 Posts	hi wacky, its now oky. my stat is at: http://www.nfsnet.org/stats2/statsre...&userid=junky& thanks for nice support ;) and i've a question, how much time is required to complete a workunit ? a workunit is composed of how many lines ? relations ? and finally, is 3 wu is "fast" for a single day ? Thanks

2004-01-25, 15:24	#7
Wacky Jun 2003 The Texas Hill Country 3²×11² Posts	Junky asks: how much time is required to complete a workunit ? a workunit is composed of how many lines ? relations ? and finally, is 3 wu is "fast" for a single day ? Well, first I should point out that in the NFSNET sysem, an assignment is not the same as a workunit. The workunit (WU) was designed to measure the computational effort in a manner that is independent from the particular sieving problem. In the case of our current project, each line is 0.00113903 WU. The number of relations generated by each line varies greatly depending on both the project and the particular line. In general the yield (relations per line) decreases with increasing line number. At very low line numbers, we were getting more than 10 relations per line. By the time we finish sieving, the current yield will have dropped to about 2/3 relations per line. Right now, the yield is a little above 1. As for how long it takes to do a WU, it all depends on your machine. A typical 1GHz PIII processor will produce about 2 WU per day if it runs full time. So is 3 WU/day fast? It certainly is not the fastest that we have seen. I have a single processor machine turning in 4.6 WU per day, and it isn't the fastest machine in the pack. However, you might note that we obtained less than 600 WU from nearly 300 machines yesterday. So 3 WU per day is well above the average. (But a number of the machines put in only 12 hours per day).

2004-01-25, 16:12	#8
junky Jan 2004 7×19 Posts	really clear explanation wacky, thanks for the info :)