Message boards :
Number crunching :
AP V7
Message board moderation
Previous · 1 . . . 5 · 6 · 7 · 8 · 9 · 10 · 11 . . . 20 · Next
Author | Message |
---|---|
qbit Send message Joined: 19 Sep 04 Posts: 630 Credit: 6,868,528 RAC: 0 |
I've crunched a few V7 tasks yesterday but I don't see much difference in speed compared to V6: http://setiathome.berkeley.edu/results.php?hostid=7327094&offset=0&show_names=0&state=0&appid=20 I use the commandline which Mike suggests for 750Ti in the readme file: -use_sleep -unroll 10 -oclFFT_plan 256 16 512 -ffa_block 12288 -ffa_block_fetch 6144 I have regular GTX750 (non Ti but factory oc'd), is that ok or are there any other suggestions? I may try to add the "tune" command soon which I already used on V6. |
Mike Send message Joined: 17 Feb 01 Posts: 34283 Credit: 79,922,639 RAC: 80 |
I've crunched a few V7 tasks yesterday but I don't see much difference in speed compared to V6: The biggest difference you will see with high blanked tasks. Tune switch might be slower on your card. With each crime and every kindness we birth our future. |
betreger Send message Joined: 29 Jun 99 Posts: 11366 Credit: 29,581,041 RAC: 66 |
A question I have is V7 doing "better" science than V6? |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
A question I have is V7 doing "better" science than V6? yes, it doing better "science". |
betreger Send message Joined: 29 Jun 99 Posts: 11366 Credit: 29,581,041 RAC: 66 |
Thanks, I eagerly await the Lunatics installer, meanwhile I will crunch resent V6s and MBs |
Urs Echternacht Send message Joined: 15 May 99 Posts: 692 Credit: 135,197,781 RAC: 211 |
I see that there is a 32-bit SSE3 version of the AP7 app for Windows (AP7 32-bit SSE3 CPU r2691), but there is none for Linux. Is that because there is no real advantage to using SSE3 instructions with Linux or is it just not posted yet? Not posted. The other versions should come up one after the other over the next two weeks. _\|/_ U r s |
Jeff Buck Send message Joined: 11 Feb 00 Posts: 1441 Credit: 148,764,870 RAC: 0 |
In the initial round of AP v7 tasks, my T7400 (running stock) received 3 CPU tasks (vs. 163 GPU tasks). The first of those is an "AstroPulse v7 v7.03 (sse2)", while the other 2 are "AstroPulse v7 v7.00". All 3 are currently running on that machine and the disparity in run times has really gotten my attention. The v7.03 (sse2) task has been running about 15 hours and is just over 77% done. If that pace continues, it should finish in about 20 hours (or less), which would be quite favorable vs. the normal AP v6 run times which tended to be in the 30-50 hour range. However, the v7.00 tasks are a very different story. The first one has been running just over 13 hours but is only about 12.6% complete, while the second one has been running about 4 hours and 20 minutes and just passed the 4% mark. If that pace is maintained, each of those tasks will take over 100 hours to complete, much worse than the comparable AP v6 tasks. I've been intermittently checking on the progress of the first two tasks for about the last 6-7 hours, and the last task since it was launched, and the progress vs. time on each of them has stayed pretty consistent with the way they look right now. So, I'm wondering, what is the fundamental difference between the v7.00 app and the v7.03 (sse2) app, and did the beta testing happen to show such a wide variation on any particular hardware setups? The CPUs on that machine are Intel(R) Xeon(R) CPU E5430 @ 2.66GHz [Family 6 Model 23 Stepping 10]. |
TBar Send message Joined: 22 May 99 Posts: 5204 Credit: 840,779,836 RAC: 2,768 |
In the initial round of AP v7 tasks, my T7400 (running stock) received 3 CPU tasks (vs. 163 GPU tasks). The first of those is an "AstroPulse v7 v7.03 (sse2)", while the other 2 are "AstroPulse v7 v7.00". All 3 are currently running on that machine and the disparity in run times has really gotten my attention. Told ya....You seem to have forgotten about AstroPulse v7 v7.00... I would tell you how to 'adjust' BOINC to run that dog with AstroPulse v7 v7.03, but, I kinda want a number of those dogs to get added to the CreditFew database. 'cause, when enough of those dogs get 'assimilated' the APv7 credits will go "to da Moon Alice...to da Moon" :-) Yes, it was observed in Beta... |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
In the initial round of AP v7 tasks, my T7400 (running stock) received 3 CPU tasks (vs. 163 GPU tasks). The first of those is an "AstroPulse v7 v7.03 (sse2)", while the other 2 are "AstroPulse v7 v7.00". All 3 are currently running on that machine and the disparity in run times has really gotten my attention. Are you looking for differences other than the SEE optimizations that the 7.03 app has? SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Jeff Buck Send message Joined: 11 Feb 00 Posts: 1441 Credit: 148,764,870 RAC: 0 |
Told ya....You seem to have forgotten about AstroPulse v7 v7.00... LOL. An ulterior motive at work here, I see. ;^) Okay, well, I'll just let them run then....but really hope I don't get any more of those bowwows. I noticed that the wingmen for both of those WUs have already reported, so I can see that they've both got 0.00% blanking which, theorectically I guess, should mean that their run times will be as fast as they possibly could be. I shudder to think what a highly blanked WU would do, even with the v7 improvements. |
Jeff Buck Send message Joined: 11 Feb 00 Posts: 1441 Credit: 148,764,870 RAC: 0 |
Are you looking for differences other than the SEE optimizations that the 7.03 app has? Well, I guess I'm really just wondering what would make a task take 5 times longer to run with one app vs. another on the same hardware. According to my notes, that CPU has SSE, SSE2, SSE3, SSSE3, and SSE4.1 instruction sets (along with a few that seem to be irrelevent). |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Told ya....You seem to have forgotten about AstroPulse v7 v7.00... Highly blanked tasks run less time now. Here are some CPU run times on one of my machines for 0% blanked & 72% blanked. SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Jeff Buck Send message Joined: 11 Feb 00 Posts: 1441 Credit: 148,764,870 RAC: 0 |
Highly blanked tasks run less time now. Here are some CPU run times on one of my machines for 0% blanked & 72% blanked. Ahh, I like the looks of that! It always seemed illogical to me that the more blanking there was, the longer the run times were. |
Josef W. Segur Send message Joined: 30 Oct 99 Posts: 4504 Credit: 1,414,761 RAC: 0 |
To put it simply, my 2.4 GHz. Core 2 Duo running WinXP 32 bit does "AstroPulse v7 v7.03 (sse)" tasks in 15 hours, "AstroPulse v7 v7.00" tasks in 70 hours (unblanked tasks in both cases). That's from many tasks at SETI Beta. The 4.67 speed ratio for 32 bit will almost certainly be more than 5 for 64 bit systems. Extrapolating from offline tests using shortened tasks, "AstroPulse v6 v6.01" would take around 26 hours on that system for those tasks. The discrepancy was noted at SETI Beta, see my post on the subject for instance. I presume the reason Eric didn't rebuild "AstroPulse v7 v7.00" is primarily that he had other things to do in limited available time which were more important. BOINC will soon have averages showing how slow that app version is, so hosts with SIMD capabilities will get tasks for it very seldom. If there are any Windows users trying to do AP tasks with hosts which don't have at least SSE, their situation becomes serious. If they were taking over half the deadline to do AP v6 work, AP v7 won't finish by the deadline. I'll probably offer builds meant for such systems at Lunatics, but we have no non-SIMD test systems so they'll effectively be Alpha in that sense. Joe |
Josef W. Segur Send message Joined: 30 Oct 99 Posts: 4504 Credit: 1,414,761 RAC: 0 |
Told ya....You seem to have forgotten about AstroPulse v7 v7.00... An 80% blanked AP v7 task run on CPU finishes in about 1/4 the time of an unblanked task. That roughly applies to all the CPU builds. Joe |
Darrell Wilcox Send message Joined: 11 Nov 99 Posts: 303 Credit: 180,954,940 RAC: 118 |
|
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
How about this one!! ;-)) AP v handled 100% blanked the same. http://setiathome.berkeley.edu/result.php?resultid=3749943302 SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Jeff Buck Send message Joined: 11 Feb 00 Posts: 1441 Credit: 148,764,870 RAC: 0 |
To put it simply, my 2.4 GHz. Core 2 Duo running WinXP 32 bit does "AstroPulse v7 v7.03 (sse)" tasks in 15 hours, "AstroPulse v7 v7.00" tasks in 70 hours (unblanked tasks in both cases). That's from many tasks at SETI Beta. The 4.67 speed ratio for 32 bit will almost certainly be more than 5 for 64 bit systems. Thanks, Joe. That's certainly consistent with what I'm seeing with just those few tasks for comparison. I hadn't seen your post on beta and TBar's earlier post on Main meant nothing to me at the time. I guess I was fortunate to get the v7 7.03 (sse2) shortly before I got the v7 7.00, so I had the two running simultaneously for comparison. Otherwise I would've been fearful that running AP v7 on the CPU was going to be a non-starter. :^) BOINC will soon have averages showing how slow that app version is, so hosts with SIMD capabilities will get tasks for it very seldom. I hope that's true, because I sure don't want to have to process 11 of the v7 7.00 tasks before the scheduler decides that's a tremendous waste of resources. In that same amount of time, I could run through 55 of the v7 7.03 (sse2) tasks! |
qbit Send message Joined: 19 Sep 04 Posts: 630 Credit: 6,868,528 RAC: 0 |
I've crunched a few V7 tasks yesterday but I don't see much difference in speed compared to V6: Thx Mike! |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
Yep, it's the only hope remains (to me personally) to get around CreditScrew. To have such slow-ugly base app released and being used as credits "normalizator". Some say even that will not help. Will see. The bright side is that when your host completes 11 tasks per each eligible app server will send almost only fastest one to it. The dark side - it can happen that host will never complete 11 tasks for its faster app and will stuck forever with slowest one. So, advise: look to your number of completed (in app state page) tasks closely. And if you would see v7.00 "plain" reaches 11 completed BEFORE any of "SSE" builds done - start to abort v7.00 tasks. But don't even think about such abortion before ;) |
©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.