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 : Problems and Help : How to only download one WU at a time? (Message 131608)
Posted 209 days ago by zombie67 [MM]Project donor
28 threads is excessive: 79353.39 seconds of total CPU time for a runtime of 3374 seconds, so you have an effective CPU efficiency of 4.25%.
Using 28 threads on one task wastes over 95% of your CPU power.
You could use a lot less threads and have about the same runtime on the PPS task,
and use the other threads to run more PPS tasks simultaneously. Your throughput will be much higher.

With 4 threads per task for PPS they each should still run in around 3400 seconds.
Except that you can run 7 of them simultaneously with -t 4 if you set your %CPU to just over 1/7 => 15%
and then your CPU efficiency will be closer to 50%, and you will have many more tasks "first" each day, while you are racking up credit 7 times faster.


Yes. Of course. I was using a single task to simplify the experiment, and to remove confusion, not to maximize crunching. ;)
2) Message boards : Problems and Help : How to only download one WU at a time? (Message 131569)
Posted 211 days ago by zombie67 [MM]Project donor
AND it ran the wu's using all 28 threads too?


Yep. Here is a task with bionic prefs set to 100%:

https://www.primegrid.com/result.php?resultid=1015618473

Here is a task set to 3.57%:

https://www.primegrid.com/result.php?resultid=1015629384

Both similar run times. Both show 28 threads in the Stderr output.

Also, both were reported first. ;)
3) Message boards : Problems and Help : How to only download one WU at a time? (Message 131562)
Posted 212 days ago by zombie67 [MM]Project donor
Next, I will try the method suggested of setting the client preferences to use only of the total threads (expressed as a percentage).


It works! It shouldn't, because the client is ignoring the CPU limitation set. But whatever, I'll take it. Thanks!

30383 PrimeGrid 7/29/2019 3:54:02 PM work fetch resumed by user 30384 PrimeGrid 7/29/2019 3:54:02 PM [sched_op] Starting scheduler request 30385 PrimeGrid 7/29/2019 3:54:03 PM Sending scheduler request: To fetch work. 30386 PrimeGrid 7/29/2019 3:54:03 PM Requesting new tasks for CPU 30387 PrimeGrid 7/29/2019 3:54:03 PM [sched_op] CPU work request: 1.00 seconds; 1.00 devices 30388 PrimeGrid 7/29/2019 3:54:03 PM [sched_op] NVIDIA GPU work request: 0.00 seconds; 0.00 devices 30389 PrimeGrid 7/29/2019 3:54:05 PM Scheduler request completed: got 1 new tasks 30390 PrimeGrid 7/29/2019 3:54:05 PM [sched_op] Server version 713 30391 PrimeGrid 7/29/2019 3:54:05 PM Project requested delay of 7 seconds 30392 PrimeGrid 7/29/2019 3:54:05 PM [sched_op] estimated total CPU task duration: 8303 seconds 30393 PrimeGrid 7/29/2019 3:54:05 PM [sched_op] estimated total NVIDIA GPU task duration: 0 seconds 30394 PrimeGrid 7/29/2019 3:54:05 PM [sched_op] Deferring communication for 00:00:07 30395 PrimeGrid 7/29/2019 3:54:05 PM [sched_op] Reason: requested by project 30396 PrimeGrid 7/29/2019 3:54:07 PM Started download of llrPPS_322531826 30397 PrimeGrid 7/29/2019 3:54:08 PM Finished download of llrPPS_322531826 30398 PrimeGrid 7/29/2019 3:54:08 PM Starting task llrPPS_322531826_1 31466 PrimeGrid 7/29/2019 4:50:13 PM Computation for task llrPPS_322531826_1 finished 31467 PrimeGrid 7/29/2019 4:50:15 PM Started upload of llrPPS_322531826_1_r1958345385_0 31468 PrimeGrid 7/29/2019 4:50:17 PM Finished upload of llrPPS_322531826_1_r1958345385_0 31469 PrimeGrid 7/29/2019 4:50:17 PM [sched_op] Starting scheduler request 31470 PrimeGrid 7/29/2019 4:50:17 PM Sending scheduler request: To report completed tasks. 31471 PrimeGrid 7/29/2019 4:50:17 PM Reporting 1 completed tasks 31472 PrimeGrid 7/29/2019 4:50:17 PM Requesting new tasks for CPU 31473 PrimeGrid 7/29/2019 4:50:17 PM [sched_op] CPU work request: 1.00 seconds; 1.00 devices 31474 PrimeGrid 7/29/2019 4:50:17 PM [sched_op] NVIDIA GPU work request: 0.00 seconds; 0.00 devices 31475 PrimeGrid 7/29/2019 4:50:19 PM Scheduler request completed: got 1 new tasks 31476 PrimeGrid 7/29/2019 4:50:19 PM [sched_op] Server version 713 31477 PrimeGrid 7/29/2019 4:50:19 PM Project requested delay of 7 seconds 31478 PrimeGrid 7/29/2019 4:50:19 PM [sched_op] estimated total CPU task duration: 7828 seconds 31479 PrimeGrid 7/29/2019 4:50:19 PM [sched_op] estimated total NVIDIA GPU task duration: 0 seconds 31480 PrimeGrid 7/29/2019 4:50:19 PM [sched_op] handle_scheduler_reply(): got ack for task llrPPS_322531826_1 31481 PrimeGrid 7/29/2019 4:50:19 PM [sched_op] Deferring communication for 00:00:07 31482 PrimeGrid 7/29/2019 4:50:19 PM [sched_op] Reason: requested by project 31483 PrimeGrid 7/29/2019 4:50:21 PM Started download of llrPPS_322532027 31484 PrimeGrid 7/29/2019 4:50:22 PM Finished download of llrPPS_322532027 31485 PrimeGrid 7/29/2019 4:50:22 PM Starting task llrPPS_322532027_1

4) Message boards : Problems and Help : How to only download one WU at a time? (Message 131561)
Posted 212 days ago by zombie67 [MM]Project donor
I ran a test and have confirmed the scheduler asks for a task per thread, even for subsequent requests after the first batch. Here is the log (28 thread machine):

26971 PrimeGrid 7/29/2019 1:17:12 PM work fetch resumed by user 26972 PrimeGrid 7/29/2019 1:17:13 PM [sched_op] Starting scheduler request 26973 PrimeGrid 7/29/2019 1:17:13 PM Sending scheduler request: To fetch work. 26974 PrimeGrid 7/29/2019 1:17:13 PM Requesting new tasks for CPU 26975 PrimeGrid 7/29/2019 1:17:13 PM [sched_op] CPU work request: 1.00 seconds; 28.00 devices 26976 PrimeGrid 7/29/2019 1:17:13 PM [sched_op] NVIDIA GPU work request: 0.00 seconds; 0.00 devices 26977 PrimeGrid 7/29/2019 1:17:16 PM Scheduler request completed: got 28 new tasks 26978 PrimeGrid 7/29/2019 1:17:16 PM [sched_op] Server version 713 26979 PrimeGrid 7/29/2019 1:17:16 PM Project requested delay of 7 seconds 26980 PrimeGrid 7/29/2019 1:17:16 PM [sched_op] estimated total CPU task duration: 262296 seconds 26981 PrimeGrid 7/29/2019 1:17:16 PM [sched_op] estimated total NVIDIA GPU task duration: 0 seconds 26982 PrimeGrid 7/29/2019 1:17:16 PM [sched_op] Deferring communication for 00:00:07 26983 PrimeGrid 7/29/2019 1:17:16 PM [sched_op] Reason: requested by project […] 28329 PrimeGrid 7/29/2019 2:17:53 PM Computation for task llrPPS_322531144_0 finished 28330 PrimeGrid 7/29/2019 2:17:55 PM Started upload of llrPPS_322531144_0_r1537324993_0 28331 PrimeGrid 7/29/2019 2:17:55 PM [sched_op] Starting scheduler request 28332 PrimeGrid 7/29/2019 2:17:55 PM Sending scheduler request: To fetch work. 28333 PrimeGrid 7/29/2019 2:17:55 PM Requesting new tasks for CPU 28334 PrimeGrid 7/29/2019 2:17:55 PM [sched_op] CPU work request: 1.00 seconds; 28.00 devices 28335 PrimeGrid 7/29/2019 2:17:55 PM [sched_op] NVIDIA GPU work request: 0.00 seconds; 0.00 devices 28336 PrimeGrid 7/29/2019 2:17:56 PM Finished upload of llrPPS_322531144_0_r1537324993_0 28337 PrimeGrid 7/29/2019 2:17:57 PM Scheduler request completed: got 28 new tasks 28338 PrimeGrid 7/29/2019 2:17:57 PM [sched_op] Server version 713 28339 PrimeGrid 7/29/2019 2:17:57 PM Project requested delay of 7 seconds


Next, I will try the method suggested of setting the client preferences to use only of the total threads (expressed as a percentage).
5) Message boards : Problems and Help : How to only download one WU at a time? (Message 131548)
Posted 212 days ago by zombie67 [MM]Project donor
As far as I know, it doesn't download 6 the second time, just the first time. After the first 6 are gone its just one at a time.


I don't think so. The BOINC scheduling code will tell the client ask for at least one task per idle thread. And since all 6 threads are idle at this point, it will ask for 6.
6) Message boards : Problems and Help : How to only download one WU at a time? (Message 131547)
Posted 212 days ago by zombie67 [MM]Project donor

|---| |---| |---| |---| |---| |---|

vs.

|----------------------| |----------------------| |----------------------| |----------------------| |----------------------| |----------------------|


I too would like to find a way for the client to download only one task when in MT mode with zero resource share, so that each task us returned as fast as possible, and the best possibility of being first. As far as I can tell, this is not possible.

|---| |---| |---| |---| |---| |---|

Calculations on the same cores cannot physically run at the same time, they are just going to be waiting for their time on the CPU core (AVX unit or elsewhere) in a queue (CPU time scheduling + fighting for CPU cache space). So yes, this is impossible and kinda weird why would you imagine it being possible. Multithreading is using 6x calculations speed.

Here's a revision of your diagram

|666| |666| |666| |666| |666| |666|

vs.

|1111111111111111111111| |1111111111111111111111| |1111111111111111111111| |1111111111111111111111| |1111111111111111111111| |1111111111111111111111|


You cannot go more than 6 at the same time verically.


I don't understand what you are saying. Your chart looks just like mine. I am comparing running a single MT task using all 6 threads, sequentially until all 6 complete, vs six single-thread tasks concurrently. They take roughly the same amount of time for all 6 to complete, either method. However, the MT version allows some to be returned earlier than others.

7) Message boards : Problems and Help : How to only download one WU at a time? (Message 131535)
Posted 212 days ago by zombie67 [MM]Project donor
When you run out of work BOINC will download one task for each core regardless of what you have for an app_config


Right. Using the OP setup with 6 threads, and running tasks MT using all 6 threads, and zero resource share, it will download 6 tasks. Then run each, one at a time until all gone. Then will download 6 more.

So the first task will be returned much faster than running it on a single thread. But the 6th task will be returned later, about the same time as if it had been run as a single thread. The end result is that MT helps return tasks faster, with diminishing advantage as the batch of downloaded tasks complete.

|---| |---| |---| |---| |---| |---|

vs.

|----------------------| |----------------------| |----------------------| |----------------------| |----------------------| |----------------------|


I too would like to find a way for the client to download only one task when in MT mode with zero resource share, so that each task us returned as fast as possible, and the best possibility of being first. As far as I can tell, this is not possible.

|---| |---| |---| |---| |---| |---|
8) Message boards : Number crunching : Max FFT size for each LLR project? (Message 131001)
Posted 231 days ago by zombie67 [MM]Project donor
Wow mackerel, thank you so much for that.


Yes! Thanks!
9) Message boards : News : AVX-512 Now Supported by LLR (Message 131000)
Posted 231 days ago by zombie67 [MM]Project donor
The expectation is that FMA3 performance of Zen 2 is comparable to recent Intel consumer CPUs, if not limited elsewhere. I've installed my CPU and running tests on it now.


This is exactly what I am asking, and you are working on the answer. Sweet!

Also, thanks for doing this!
10) Message boards : News : AVX-512 Now Supported by LLR (Message 130985)
Posted 232 days ago by zombie67 [MM]Project donor
Does this mean the Ryzen 3000 are now comparable to Intel at FMA3? It is truly fixed now, and performs about the same as intel?

I notice the intel X processors are much faster, but I assume that is due to using AVX-512.


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2020 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 0.00, 0.00, 0.00
Generated 27 Feb 2020 | 0:07:06 UTC