PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
1) Message boards : Number crunching : No respect of Buffer sizes with Mac OS X (Message 83882)
Posted 2054 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
Knowing this problem, why do you always try to shorten the deadlines? The shorter they are, the smaller the buffers should be, impacting all projects. Reducing the buffers to 0.01 or 0 just because you run PrimeGrid and that there's a bug somewhere with PG WUs is a nonsense.

Look: I received my last bunch of GPU WUs on Monday. All finished tasks have been cleared as soon as completed, but I still have more than 100 hours of estimated work in queue (when the buffer maximum size is 0.5d = 12h). As I've stopped running all PG CPU subprojects, the estimation is quite good.
They'll never complete in time. May I suggest that PG Team fixes the bug? A good challenge, isn't it.
2) Message boards : Number crunching : No respect of Buffer sizes with Mac OS X (Message 83857)
Posted 2056 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
I'm running BOINC 7.4.36 on an iMac with an AMD GPU and OS X 10.10.2.

When receiving work from Primegrid, it gets a bunch of WUs. Their total estimated time is up to 10-12 times the maximum size of the Input buffer for GPU tasks, plus up to 10-12 times the maximum size of my Input buffer for CPU tasks (according to the number of threats available). Completed WUs are cleaned regularly. When going down to 0, each queue is refilled the same way.

Is there any reason for that?

Please note that I'm running several BOINC projects, and that Primegrid is the only project that causes this problem.
3) Message boards : The Riesel Problem : Reducing TRP-Sieve deadline to 3 days, or increase everything to 4 days? (Message 83838)
Posted 2057 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
In my computers, the cleaning is immediate but the execution isn't (due to waiting list and time sharing). I don't bother about deadlines, except when they become critical to empty the waiting list in time. It's already the case for a lot of "short" Primegrid subprojects. A week-end may last up to 110 hours, and running more than 8h/threat or 8h/GPU in each available working day is totally irrealistic in a timesharing environment.
4) Message boards : The Riesel Problem : Reducing TRP-Sieve deadline to 3 days, or increase everything to 4 days? (Message 83837)
Posted 2057 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
I have the same feeling. I'm in trouble using Primegrid on all my computers as the deadlines are too short. But there may be another problem behind.

I'm running openclPPSsieveMAC on an iMac / AMD GPU / BOINC 7.4.36, and no other project or subproject uses the GPU. My working conditions require a reasonable buffer size (6 to 12 h of work).
Then the deadline is a problem because it receives a lot of openclPPSsieveMAC WUs at each Primegrid download, representing 10 to 12 times the maximum buffer size. There is no problem for cleaning: it's immediate when a WU complete.

Is there anybody able to explain why ?
5) Message boards : The Riesel Problem : Reducing TRP-Sieve deadline to 3 days, or increase everything to 4 days? (Message 83735)
Posted 2062 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
A wise solution should be to multiply all deadlines by a factor of 2 or 3.
6) Message boards : The Riesel Problem : Reducing TRP-Sieve deadline to 3 days, or increase everything to 4 days? (Message 83733)
Posted 2062 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
You're right: in order to get 0.5+0.5, I've set buffer sizes to 0,5d/1d.
Yesterday, I tried to go down to 0/0. The queues have been reduced to the minimum, but, as some project servers (and my Internet connection) are not 100% available, I spent long times without work. So I absolutely need a reserve of work.
I went back to 0.1d/0.2d. With 3 CPU threats active, I should have between 7,5h and 15h of total CPU load and between 2,5h and 5h of GPU load. Please note that I'm running 2 projects at 100-100 and that the GPU don't work when the PC is used.
FALSE: after an automatic update, the total CPU workload is now 164h for Primegrid (7 WUs for 4 subprojects) + 3 hours for another project (6 WUs), and the GPU workload is 25h for a single Primegrid subproject (6 WUs in queue + 1 in progress).
Running 24/7 in absence of other use, the maximum available time per project and per day for my PC is 24h x 3 threats x 0.5 = 36h, and 164h correspond to 4.5 days.
With 4-day deadlines, it's already hard to complete in time. With 3-day deadlines, I'm completely out.
7) Message boards : The Riesel Problem : Reducing TRP-Sieve deadline to 3 days, or increase everything to 4 days? (Message 83720)
Posted 2062 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
I was used to work with the same figures (0.5 day + 0.5 day), and I usually have around 1 day of work in queue for each project or subproject. When a queue decreases to its lower limit, I'm receiving a bunch of new WUs. But Primegrid behaviour is very strange : GPU tasks duration is underestimated whilst CPU tasks duration is overestimated, so the planned execution time for both is playing yoyo in a 1 to 4 range. When high, I have only a few WUs in queue. When low, I'm receiving a lot of WUs, so they turn to high priority, monopolize the resources and fall out of time. A real mess! The only solution I've found is to reduce the buffer size: I'm now at 0.2d + 0.2d and it runs. Using lower values is risky as I don't have a permanent Internet link to all Project servers.
You should understand that, as soon as computers have several WUs in queue, there is no relation between deadlines and task duration. If PRIMEGRID reduces its deadlines, participants will have to reduce their pending workload, that will effect ALL other projects.
I will try to go to 0d+0d. If not satisfactory, I will stop crunching for Primegrid.

Please note that I'm only running short tasks, as a single PrimeGrid long task require more than 50% of resources availability of 1 thread to complete in time.
As example, an openCLPPSsleeve WU takes 4 hours to complete on my IMAC, and I have 1 WU running and 8 WUs in queue. Total working time: 36h. For heating protection, my GPU doesn't work when I'm using the computer. So I should clear this queue in around 3 days if running only this subproject. But I'm using BOINC, based on multi project distributed computing. If running only 2 projects and 2 Primegrid subprojects in Parallel with GPU WUs, only 25% of GPU resources are available for this single subproject, so it will not take 3 days but 12 to clear the queue and get new WUs. Look at the deadline! This has no sense in a multi project environment, except if Primegrid plans to monopolize 100% of participants resources for its own project.

The Primegrid deadlines already are ridiculously short. Don't try to reduce them any more. The most important Challenge for this project is to make participants' life easy.
8) Message boards : The Riesel Problem : Reducing TRP-Sieve deadline to 3 days, or increase everything to 4 days? (Message 83691)
Posted 2063 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
I agree that the 1-day idea is a little sarcastic, but running your project gives headaches.

Please have a look on the participants' point of view : I really think that you misunderstand the way BOINC application works on THEIR computers.

Suppose that I have a 1 day buffer and 2 cores active. I will roughly receive WUs for a total of 48h workload. If the WU duration is 2 hours, I will receive 24 WUs. If it's only 1 hour, I will receive 48 WUs. As the 48 new WUs will complete in the same time than the 24 old WUs, there is absolutely no reason to reduce the deadline.

Reducing the deadline as you do in your project has an unique effect : force all WUs to run in high priority, with the risk of not completing in time. May be it's your goal, but let me tell you that this is a nonsense in a time-sharing environment!
9) Message boards : The Riesel Problem : Reducing TRP-Sieve deadline to 3 days, or increase everything to 4 days? (Message 83686)
Posted 2063 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
Following your theory, why don't you reduce the deadline to 1 day ?
The question is not the task duration but the availability of computers for this particular application. The shorter is the task, the higher is the number of WU received so the total workload received in one batch remains the same. Reducing the deadline of short tasks is a nonsense : it only makes participants' work very difficult, as they should drastically reduce the work buffer size used by the BOINC application for ALL projects and select computer by computer the PrimeGrid subprojects to be run if they want to complete the WUs in time.
Other point : I don't know how long is the week-end break in your country, but in mine, it's around 60 hours, so it will be very difficult to complete a 3-day (72h) deadline.
You must consider that a lot of computers running BOINC don't work 24/7 and aren't 100 % available for an unique subproject.
10) Message boards : Problems and Help : Problem with runtime increase on CPU and GPU (iMAC) (Message 80832)
Posted 2168 days ago by [AF>France>Aquitaine>Cote-Adour-et-Gaves]Bernard 64250
No progress on this subject after 1 year : CPU tasks durations are still overestimated by an x10 factor when GPU tasks ones are acceptable. If you run both types on the same computer, it's a real mess, but apparently nobody cares !


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2020 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 0.84, 1.25, 1.12
Generated 25 Oct 2020 | 22:14:41 UTC