| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Community
Forum: Hardware Chat Room
Thread: Suddenly bad performance on Athlon 200GE with MCM. |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 36
|
|
| Author |
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
So what to do to still have the 'user' benefit of hyperthreading, but speed up the MCM/BOINC crunching: Enable hyperthreading and set BOINC to 50% of CPU threads. The OS is smart enough to merge the power of the 2 threads, or rather, not split the cores for BOINC when the device is purely crunching and when you 'use' the device it will siphon off a little juice for your personal employ.
----------------------------------------P.S. The science engine was not changed between the previous and current target as far as I know, plus the size of the jobs was actually reduced to accommodate that crunching is now also taking place on Android (see post tech armstrdj). [Edit 1 times, last edit by Former Member at May 11, 2020 4:56:44 PM] |
||
|
|
floyd
Cruncher Joined: May 28, 2016 Post Count: 47 Status: Offline Project Badges: 
|
Hyperthreading on this CPU leads to extremely bad and varying performance with the Sarcoma work units of Mapping Cancer Markers. That's not what I expected. The MIP code is known to need much L3 cache, even larger processors can only run few of those tasks without degrading performance. I'd be surprised to see your little thing run four at full speed. Or do you run mixed workloads? That would be a good explanation for varying performance.The problem does not exist with Micriobiome Immunity work units When I disable hyperthreading, I get very good run times of MCM Sarcoma between 1h37mins up to 2h07mins. Actually two tasks in 150 minutes versus one in 97 minutes is a nice gain. You can't expect much more. 7 hours is too long though. Does that happen often?When hyperthreading is ENABLED though, I get run times between 2h30mins up to 7 hours! MCM tasks take between 2.57 and 3.4 hours on my Ryzen 1700 with SMT. Judging by the numbers that processor can be compared to yours, it's just four times the size. Except ... yours has integrated graphics. Do you use that? I read it costs quite some CPU power. |
||
|
|
Piri1974
Cruncher Joined: Jun 3, 2018 Post Count: 24 Status: Offline Project Badges:
|
So what to do to still have the 'user' benefit of hyperthreading, but speed up the MCM/BOINC crunching: Enable hyperthreading and set BOINC to 50% of CPU threads. The OS is smart enough to merge the power of the 2 threads, or rather, not split the cores for BOINC when the device is purely crunching and when you 'use' the device it will siphon off a little juice for your personal employ. P.S. The science engine was not changed between the previous and current target as far as I know, plus the size of the jobs was actually reduced to accommodate that crunching is now also taking place on Android (see post tech armstrdj). Thanks a lot for the info. I'll set BOINC to use only 2 threads and see if that helps. But I NEVER use this pc for something else than BOINC, not even Word. Could it be that the Sarcoma WUs need more cache memory? As you can see on the product page of the Athlon 200GE, it has 1Mb of L2 cache and 4Mb of L3 cache, which is the least of all 1st Gen Ryzen CPUs. For example: One step up you have the Ryzen 3 1200 which has already 2Mb L2 an 8Mb L3 (but no hyperthreading though). The first one with hyperthreading is the Ryzen 5 1600, which has 3M L2 and a whopping 16Mb L3. Could it be that the difference in cache setup causes that only the small Athlon has this problem? But now that I think of it: My slowest cpu running Sarcoma (and nothing else) is an AMD A8-9600. That cpu has only 2Mb of L2 and no L3 at all. That cpu has no hyperthreading, and does Sarcoma WUs in around 3 hours, +- 15 minutes. So quite constant performance. We all agree that the performance degradation I am seeing is not normal with hyperthreading: virtual cores do bring a small performance deficit per core, but never a 200% performance loss. [Edit 6 times, last edit by Piri1974 at May 11, 2020 7:28:18 PM] |
||
|
|
Piri1974
Cruncher Joined: Jun 3, 2018 Post Count: 24 Status: Offline Project Badges:
|
The MIP code is known to need much L3 cache, even larger processors can only run few of those tasks without degrading performance. I'd be surprised to see your little thing run four at full speed. Or do you run mixed workloads? That would be a good explanation for varying performance. Thanks for the ideas. I never run mixed workloads on this PC. It only runs MCM. I just switched to MIP for some WUs because someone advised me to check if the problem persists with MIP WUs. And that happened to be a good test, because no, with MIP I don't see the problem, not even with 4 parallel threads. Actually two tasks in 150 minutes versus one in 97 minutes is a nice gain. You can't expect much more. 7 hours is too long though. Does that happen often? ... yours has integrated graphics. Do you use that? I read it costs quite some CPU power. Let's say that 1 in 4 WUs takes longer than 6 hours. But I also have some WUs which are done in 2h30mins. A normal run time per WU on this cpu should be a little slower than with hyperthreading disabled, so I guess a 'normal' run time should be about 2h to 2h30mins I guess. Yes I have integrated graphics and I am currently using it indeed. But until recently I had a dedicated entry level GT730 graphics card in it. That did not seem to change anything though, but I'll put it back in when I find the time to double check. I use that card for GPUGrid from time to time, but then I'm not running anything else. Like I said, I don't do mixed workloads on this machine. But I never use this PC for anything else than BOINC, I don't even run Word on it. So I guess even the integrated graphics is mainly sitting there, doing nothing. To compare: My slowest CPU running Sarcoma WUs is an AMD A8-9600. 4-core, 4 thread with just 2Mb of L2 and no L3 at all. So that cpu has no hyperthreading, and does Sarcoma WUs in around 3 hours, +- 15 minutes. So quite constant performance. :-) [Edit 2 times, last edit by Piri1974 at May 11, 2020 7:31:37 PM] |
||
|
|
Piri1974
Cruncher Joined: Jun 3, 2018 Post Count: 24 Status: Offline Project Badges:
|
Anyway, what I will do tomorrow and the day after is this:
Activate hyperthreading again, but let BOINC only use 50% of the cores. So we are certain that it will still use only the 2 physical cores, am I right? If the performance remains good, I will go 1 step further and let BOINC use 75% of the cores. This means it will need to use 1 virtual core. I am looking forward to the result of that exercise. I will keep you informed. Again many thanks for all input, help and ideas. Carl Philip |
||
|
|
floyd
Cruncher Joined: May 28, 2016 Post Count: 47 Status: Offline Project Badges:
|
A normal run time per WU on this cpu should be a little slower than with hyperthreading disabled Not in my experience. A 25% drop in speed is the best I've seen, that makes a 50% overall gain when hyperthreading. The worst I remember was 10% gain but other people claim they even lose performance. It all depends on the applications. Again, your numbers look good to me except for the occasional long-runners. |
||
|
|
|