Message boards :
Number crunching :
credit equals time?
Message board moderation
Author | Message |
---|---|
Garry Send message Joined: 7 Jul 02 Posts: 40 Credit: 535,102 RAC: 1 |
background: I recently had problems with too-frequent late work units. Good info at https://setiathome.berkeley.edu/forum_thread.php?id=84954. Seems fixed. goal: Balance time among SETI, Rosetta, and Einstein. current: 1. All projects have equal priority. 2. Statistics tab shows credit ratios approx 30:10:3 for Einstein:Rosetta:SETI. 3. The scheduler seems prone to allow Einstein to aggressively send work units, all of them roughly 24 hours long. Rosetta gets in fairly often with work units of roughly 6 hours (by adjusting a parameter on their project settings). When SETI gets in, they might get four units of 3 hours in. 4. This minute: All tasks running are Einstein. â“ Credit per time varies among projects. Presuming time is indeed equal, are these ratios typical? (Seems extreme.) â“ Presuming these ratios reflect time, is there a better path than cutting Einstein priority to 1/3 of present and increaing SETI priority to a bit over 3 times present (targeting 10:10:10 ratios on the statistics tab)? Thanks in advance for any answers! |
Mr. Kevvy Send message Joined: 15 May 99 Posts: 3799 Credit: 1,114,826,392 RAC: 3,319 |
Two things that I found help with work overload having Einstein@Home as a backup project: 1) Set its Resource Share to zero. Then it will only download enough work to keep each CPU core or GPU busy without keeping a cache (this doesn't apply for anyone "spoofing" their GPU count of course.) 2) Disable CPU work. As you noted they take too long to complete compared to the GPU work (and add little credit as CPU is so slow in comparison.) Both of these settings are in the project's preferences in your Einstein@Home account page. |
rob smith Send message Joined: 7 Mar 03 Posts: 22491 Credit: 416,307,556 RAC: 380 |
2. Statistics tab shows credit ratios approx 30:10:3 for Einstein:Rosetta:SETI. Your observation of the credit awarded is fairly typical - estimates do vary, but typically Einstein is between 10 and 15 times that of SETI with Rosetta somewhere in between. (If you really want to chase credit try Colatz) â“ Credit per time varies among projects. Presuming time is indeed equal, are these ratios typical? (Seems extreme.) Each project is at liberty to award credit as they feel fit, thus using credit to compare between projects is difficult, if not impossible. SETI uses a credit system called "Credit New", which is supposed to award credit proportional to the number of "FLOPS" that a task has taken. Many other projects do not use CreditNew, but either use a fixed credit per task, or some variation on the credit per task duration. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Garry Send message Joined: 7 Jul 02 Posts: 40 Credit: 535,102 RAC: 1 |
Valuable answers. Thanks. Follow-on questions, because I just noticed BOINC Manager > Projects tab > Avg. work done column. What units does the Avg. work done column present? Something like minutes per day, maybe? Or credit per day the project awarded? If it's not obvious from the above answer: Are the ratios there representative of the way the projects are sharing the processor? This minute, my ratios are roughly 13:6:2 for Einstein:Rosetta:SETI. Thanks in advance. |
rob smith Send message Joined: 7 Mar 03 Posts: 22491 Credit: 416,307,556 RAC: 380 |
It is a rolling average of the daily work done, with a half-life of about ten days. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14676 Credit: 200,643,578 RAC: 874 |
It is a rolling average of the daily work done, with a half-life of about ten days.'Work done' being measured by the credit awarded - which is not a very good measure. Disbelieve anybody who says that credit can be converted back to derive the number of FLOPs that earned it, according to the cobblestone scale. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13841 Credit: 208,696,464 RAC: 304 |
Disbelieve anybody who says that credit can be converted back to derive the number of FLOPs that earned it, according to the cobblestone scale.... because Credit New doesn't actually award Credit according to the definition of a Cobblestone. Grant Darwin NT |
Garry Send message Joined: 7 Jul 02 Posts: 40 Credit: 535,102 RAC: 1 |
Again, much appreciated answers. 🙠Based on what I now know, if I want to balance processor time among my three projects, I want settings that give each project an equal average number of processors in use. The best way I know to observe that is manually. Potential avenues of further investigation: Coding something with the BOINC command line. Researching options in the BOINC XML files. I'll see what I can do without those first. Have great days everyone! ... |
Kissagogo27 Send message Joined: 6 Nov 99 Posts: 716 Credit: 8,032,827 RAC: 62 |
one method to do https://boinc.berkeley.edu/trac/wiki/ClientAppConfig |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13841 Credit: 208,696,464 RAC: 304 |
if I want to balance processor time among my three projects, I want settings that give each project an equal average number of processors in use. The best way I know to observe that is manually.Then you are going to be very busy & very frustrated as you fight the BOINC Manager as it tries to honour your resource share settings, and you actively work against it's efforts. Grant Darwin NT |
Garry Send message Joined: 7 Jul 02 Posts: 40 Credit: 535,102 RAC: 1 |
An update on my efforts to balance processor usage, in case it helps someone: All ratios are in the form SETI:Rosetta:Einstein. THE SHORT VERSION Current experience is
<app_config> <project_max_concurrent>1</project_max_concurrent> </app_config> Work credit is not a perfect metric for this purpose. I don't know of anything else to try. I'm eager to see whether it is "good enough". The SETI resource share of 3000 is recent. My previous setting was 2000; maybe the scheduler will accept more SETI tasks. The Einstein resource share of 300 is recent. The previous setting was 0. 300 is 10:1 with the SETI resource share. Maybe a good starting point. As to Einstein's resource share of 0: I confirmed @Mr. Kevvy's info below: That's different than "no new tasks" (umm ... really? Resource share of 0 doesn't mean "no resources"? 🤔 Maybe that's because "resource share of 0" should be different than "not contributing". Curious.) @Mr. Kevvy: Thanks. I believed you. And I wanted to see what it did. Maybe not useful for my goals. Results included acceptance of an out-sized group of Einstein tasks; the current group. The Einstein concurrent task limit of 1 is recent. I have recently had 3, then 2, responding to completing the Einstein tasks accepted. Maybe my use of this limit has lowered the risk of accepting another out-sized group of tasks. Maybe the concurrent task limits are only useful for speeding convergence of scheduling data. It'd be nice to be able to relax them. Given the 10-day scheduler data half-life (thanks @robsmith), half of all prior decisions will be out of the scheduler data in 10 days, and half again (a total of 75%) in 20 days. Maybe that's a good point to next assess current settings. As to all the thanks I owed here: It's great to have a vibrant community. 🀠|
Garry Send message Joined: 7 Jul 02 Posts: 40 Credit: 535,102 RAC: 1 |
Another update on my continuing efforts to equally balance time among SETI, Rosetta, and Einstein. In case it helps someone. THE SHORT VERSION Current experience is: - Changes: -- BOINC Manager > Options > Computing Preferences (see below) --- Use at most 50% of the CPUs. (My computer has eight threads.) --- Use at most 100% of the CPU time. -- Number of concurrent tasks: unrestricted for all projects. - Unchanged: (Ratios expressed in form SETI:Rosetta:Einstein.) -- BOINC Manager > Options > Computing preferences --- Store at least 0.04 days of work (approximately 1 hour) --- Store up to an additional 0.01 days of work (approximately 15 minutes) -- Rosetta web site > project preferences > Target CPU run time 6 hours -- Resource share 30:10:3 -- Scheduler data half-life: 10 days (default) -- Switch tasks each 360 minutes (6 hours). Early results: - The switch from "use 100% of CPUs" to "use 50% of CPUs" resulted in my computer running half as many BOINC tasks, but all threads of the computer remaining at nearly 100% busy. More detail below. - The switch from "use 100% of CPUs" to "use 50% of CPUs" resulted in quickly reducing task inventory from something 9-15 tasks to as low as 4 tasks. - SETI tasks are arriving at this computer in such small volume, it is impossible to balance time among the three projects. - When SETI tasks arrive, they get promptly process and exit the system. If only they'd get prompt replacement ... MORE DETAIL I downloaded "Core Temp 1.15.1" (Windows) to monitor CPU temperatures. I have no reason to doubt it operates as advertised. It reported periods of excessive temperature. That motivated me to switch configuration to "use 50% of CPUs". This package has an "overheat protection" feature. Among the possible responses, it can run a batch file if it senses overheat. I wrote one that issues a command to the BOINC command program to top all BOINC activity. I set it for 85 deg C; it activated once, but only then. I tried TThrottle, advertised to throttle BOINC in response to temperature. It may not have received maintenance in a couple years. It starts operation by running a calibration routine. That routine drove my CPU temperatures near the "do not exceed" temperature for several seconds. Not welcome here. I'm confident that the scheduling function changed the mix of tasks sent to my computer when I changed to "use 50% of CPUs", but I'll have to operate this way for a while to have numbers. I think each project sends fewer tasks and shorter tasks. Maybe, the default 10-day half-life for scheduler data was too short for this computer when I was using "all CPUs". I know some will say this is BOINC heresy. If I cannot go back to "all CPUs", I cannot prove it either way with data. Maybe, the default 10-day half-life will work for this computer with "50% of CPUs". The shorter tasks, fewer number of tasks sent at a time, and less demanding deadlines may make it fine. I'm told to expect the number of tasks completed per week to decrease with "50% of CPUs", but not 50% as one might suspect. I'm interested to measure that effect. News as I have it. |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.