Message boards :
Sixtrack Application :
AVX Sixtrack version
Message board moderation
Previous · 1 · 2 · 3 · 4 · Next
Author | Message |
---|---|
Send message Joined: 27 Sep 08 Posts: 803 Credit: 649,994,523 RAC: 239,714 |
Happy to try for you in a few weeks when I can build my x299 rig |
Send message Joined: 16 Sep 17 Posts: 100 Credit: 1,618,469 RAC: 0 |
My system stopped receiving AVX tasks on Thursday, is consistently getting SSE2 instead. |
Send message Joined: 28 Sep 04 Posts: 675 Credit: 43,534,170 RAC: 15,603 |
My hosts keep swapping between the avx and sse2, just like it shows on the application details page. Which ever has higher average proeceesing rate that app is used. So I think that the speed difference between the two apps is about nonexistent. It depends a lot more of the tasks you happen to receive. |
Send message Joined: 27 Sep 08 Posts: 803 Credit: 649,994,523 RAC: 239,714 |
I think James said that there was a speed up when you run the same WU with and without AVX, since we don't get to do that test it's looks more variable on our end. I have 1 PC that seems stuck runinig SSE2, the others were running all AVX but seem to have swicthed around in the last few days. |
Send message Joined: 16 Sep 17 Posts: 100 Credit: 1,618,469 RAC: 0 |
Any chance we can re-open this issue? I've been stuck with SSE2 instead of AVX for too long. At this point I doubt this is going to remedy itself. The algorithm that is supposed to re-evaluate my system is not doing its job. It might even be the culprit, since I was getting AVX initially. |
Send message Joined: 16 Sep 17 Posts: 100 Credit: 1,618,469 RAC: 0 |
Any chance we can re-open this issue? I've been stuck with SSE2 instead of AVX for too long. At this point I doubt this is going to remedy itself. The algorithm that is supposed to re-evaluate my system is not doing its job. It might even be the culprit, since I was getting AVX initially. AVX has returned with the latest wave of SixTrack tasks! Yay, thank you! Although I might have done something to shake up the ol' rusty bits?! I was setting up an app_config.xml for another project, which escalated into creating an app_config.xml for LHC's ATLAS. I didn't even define SixTrack as an <app>. (I thought I wouldn't receive SixTrack without a definition, but I do ...) In connection with my ATLAS issues, I reset both "MAX # jobs" and "MAX # CPUs" in the website preferences to "No Limit". After reading the config files and updating my projects in the BOINC client, new SixTrack tasks had been released and AVX started pouring in. Not sure what did it though. |
Send message Joined: 18 Dec 15 Posts: 1687 Credit: 103,038,808 RAC: 126,639 |
Perhaps someone from the team could enligthen us as to what are the criterions for receiving either sse2 or avx tasks. Today, I upgraded one of my PCs to Windows 10, so the CPU which is AVX capable can in fact apply this feature. From the downloaded Sixtrack tasks after starting LHC, 5 were AVX, and 14 were sse2. How come? Why did I not receive AVX tasks only, and no sse2 tasks? It would just be interesting to understand what's behind it. |
Send message Joined: 28 Sep 04 Posts: 675 Credit: 43,534,170 RAC: 15,603 |
I think that when you first attach a host to a project, the server will select the application(s) based on features your CPU supports. Because sometimes there is a difference in speed between different CPU features (AVX/SSE2 etc.) and the "best" isn't always the fastest, several applications are selected and their output compared to determine what is fastest. You can see how different applications perform after you have finished some tasks (>10 is needed for reliable result) by viewing your Host's Details from the server page and select "Application details". All applications running in that Host are listed there with "Average processing rate". The highest GFLOPS value application is selected and will be used from there on. From time to time a set of tasks with different application are run just to verify that the fastest application is in action. The different runtimes between sixtrack tasks can make the selection biased. For instance most of my Hosts currently show SSE2 being faster than AVX. |
Send message Joined: 15 Jun 08 Posts: 2401 Credit: 225,443,980 RAC: 123,698 |
A method to run Sixtrack independent from BOINC It's tested on a linux host but may also work on windows. 1. Change to the folder /<BOINC_client_base>/projects/lhcathome.cern.ch_lhcathome/ It contains the executables, e.g. sixtrack_lin64_4630_avx.linux, and (it you got work) the job data, e.g. LHC_2015_LHC_2015_234_BOINC_errors_1.2__12__s__62.31_60.32__5.5_5.6__5__30_1_sixvf_boinc41025.zip 2. Copy the executable you want to test together with a *.zip to a fresh runtime folder. Place that folder outside the BOINC tree. 3. Rename the *.zip to "Sixin.zip" 4. Open a console window, cd to your runtime folder and run "time ./sixtrack_lin64_4630_avx.linux" 5. Wait. As usual it may take between a few seconds and a couple of hours. 6. Clean the runtime folder and repeat the test using sixtrack_lin64_4630_sse2.linux. 7. Compare the different runtimes. Your runtimes are very close together? Then you may get SSE2 for a while and then AVX for another while. Different input parameters lead to different runtimes and - as Harri already explained - this influence can be greater than the SSE2/AVX factor. |
Send message Joined: 18 Dec 15 Posts: 1687 Credit: 103,038,808 RAC: 126,639 |
Harri Liljeroos wrote: ...there is a difference in speed between different CPU features (AVX/SSE2 etc.) and the "best" isn't always the fastest.same here - so I am wondering what's supposed to be the advantage of AVX over SSE2 ??? |
Send message Joined: 28 Sep 04 Posts: 675 Credit: 43,534,170 RAC: 15,603 |
Harri Liljeroos wrote: I think the advantage depends on what kind of maths has to be done. Here the sixtrack application maybe doesn't gain so much. For example in Seti, AVX is usually faster if it is supported by the CPU |
Send message Joined: 26 Oct 18 Posts: 91 Credit: 4,188,598 RAC: 0 |
Hi, I was browsing things out there. What to think about that: https://lhcathome.cern.ch/lhcathome/workunit.php?wuid=171155568 Two linux hosts running AVX tasks. Run times for i5 host have been about dozen seconds but there's one task that run for 1300 sec. That variation looks 'natural' to me. Tasks have validated succesfully. Run times for i7 host have been 1 or 2 sec (in about thousand tasks I saw one that run for 3 sec). That such a constant speed looks a bit 'unnatural' to me. From 3000 sixtrack tasks that host got then 1500 validation inconclusives and 100 invalids and 0 valids so far. 1500 waiting for validation. Why is that ? Is something incompatible and boinc doesn't recognize that ? Would some mechanism need betterment ? |
Send message Joined: 15 Jun 08 Posts: 2401 Credit: 225,443,980 RAC: 123,698 |
Runtimes of just a few seconds usually mean that the input parameters result in an instable beam hitting the wall very early. One goal of SixTrack is to sort out those input parameters and continue with those resulting in a stable beam. Everything is fine as long as the wingcomputer of the "fast" computer confirms the short runner. From the BOINC perspective a mix of extremely short/long runtimes will lead to a confused BOINC client that miscalculates runtime estimation and credit reward. Unfortunately this can't be avoided. |
Send message Joined: 26 Oct 18 Posts: 91 Credit: 4,188,598 RAC: 0 |
Runtimes of just a few seconds usually mean that the input parameters result in an instable beam hitting the wall very early. Thanks for explaining that. I didn't know that task run time is related to how long something stays safely in the middle of that pipe until understearing at long corners and colliding into wall. Everything is fine as long as the wingcomputer of the "fast" computer confirms the short runner. Well, that's what I'm still wondering. After about one day sixtrack history currently looks like this for that faster host: State: All (3966) · In progress (5) · Validation pending (1303) · Validation inconclusive (2249) · Valid (0) · Invalid (409) · Error (0) I noticed afterwards that it has run a few sse2 tasks also. So it's not experiencing that problem with avx only... but overwhelming majority of those tasks have been avx anyway. It looks like practically all 4000 tasks sent to that host run 1-2 sec. Is it normal or statistically even likely that one computer could receive only that kind of tasks ? It is running them and doing something with them... they haven't errored out. But why all of those tasks can end up invalid ? That just makes me think something could be wrong, perhaps with that host. On the other hand, it would be a waste if a host could theoretically run 25000 sixtrack tasks in a week without getting a single succesfull validation. Does project server have somekind of special plan for this kind of situations if it continued ? |
Send message Joined: 15 Jun 08 Posts: 2401 Credit: 225,443,980 RAC: 123,698 |
The computer returning nothing but failures represents not more than 0.5% of the total computing power currently in use for SixTrack. OTOH each outlier is resend until a WU reaches a predefined #valids or #invalids/inconclusives. From the project perspective none of the results are lost. In such a case it is practically impossible to identify the reason why a single computer returns only invalids and to stop sending tasks to it while the campaign is in progress. The runtimes per task are simply too short. The project server has indeed a max #tasks per day per app version that are send out to each computer. This is currently set to 500. Since valid tasks are not counted the real number can be much higher. |
Send message Joined: 26 Oct 18 Posts: 91 Credit: 4,188,598 RAC: 0 |
The project server has indeed a max #tasks per day per app version that are send out to each computer. I'm not sure if I understood that right but I don't currently see any restriction like 'max 500 per day' slowing down that computer. Sixtrack tasks have been available for about five days now but that computer has run thru almost 13K tasks from this batch already. More than 2K per day. There are no valid tasks yet. |
Send message Joined: 13 May 14 Posts: 8 Credit: 8,408,903 RAC: 0 |
Perhaps if your CPU downclocks to the base core speed when AVX instructions are detected and if the difference between the base core speed and all core boost is great enough, then SSE2 could be faster because it doesn't cause the downclocking...I don't think. |
Send message Joined: 13 May 14 Posts: 8 Credit: 8,408,903 RAC: 0 |
Perhaps if your CPU downclocks to the base core speed when AVX instructions are detected and if the difference between the base core speed and all core boost is great enough, then SSE2 could be faster because it doesn't cause the downclocking...I don't think. I guess what I meant was the reason to keep AVX is because some systems don't downclock so much when running AVX and it makes it faster than SSE2 |
Send message Joined: 7 May 17 Posts: 5 Credit: 145,820 RAC: 0 |
Is there a way to limit receipt of SixTrack tasks to only one type (i.e., only receive AVX or only SSE2)? |
Send message Joined: 15 Jun 08 Posts: 2401 Credit: 225,443,980 RAC: 123,698 |
No, since they are app versions of the same app. It's not possible due to the design of the BOINC client. See this post in the same thread: https://lhcathome.cern.ch/lhcathome/forum_thread.php?id=4581&postid=33909 |
©2024 CERN