Message boards :
Number crunching :
Using other path...
Message board moderation
Author | Message |
---|---|
Send message Joined: 19 Jul 10 Posts: 589 Credit: 18,926,809 RAC: 4,448 |
One of my computers run out of SETI work, so I wanted to do something for MilkyWay with it. It's this one here. Yes, I know it's old. Anyway, the CPU supports SSE and BOINC recognizes it according to the client_state.xml: <p_features>fpu tsc sse 3dnow mmx</p_features> So my question is: is there no SSE path in the application or is there something going wrong? It's also running very slow... I mean, I understand, that I will not have the same performance as with the old Gipsel app (~11 hours for a 213cr WU), but the current estimate (calculated by myself from the current progress) is ~ 42 hours for a 213cr WU, that's 4 times longer, on my Laptop it was about 1.5 times longer compared to the Gipsel app. That's what in the std_err so far: <search_application> milkyway_separation 1.00 Windows x86 double </search_application> |
Send message Joined: 8 Feb 08 Posts: 261 Credit: 104,050,322 RAC: 0 |
Below the SSE2 path there is only x87, so there is nothing going wrong. I think cruncher said there wasn't really a difference between SSE and X87 with the functions he was giving to Matt. Gipsel used the library functions and the MS compiler for the X87 and SSE version because the Intel compiled versions for these older CPUs were slower (but for SSE2 and up the Intel compiled versions were running faster). From an old PM from Gipsel: - Athlon X2 CPUs with the SSE2/3 app are almost twice as fast (at same clock) as the SSE app on AthlonXPs - with SSE2 AthlonX2 shows only about half of the performance of a Phenom or Core2 if the code is vectorized properly (otherwise there would be no difference) Hope that gives you some idea about the speed differences between the different version and cpus, even that the above is for the old and outdated apps. I think Matt is using an actual MS compiler, but with other libraries and some important functions replaced by functions from cruncher. Started mw on an old MP2200 again to see what runtimes I get with the actual app on it. As expected, the stderr shows 'other path'. With less than 1% done it looks like 50hrs with mw using around 80% cpu and 20% used by other progs running. So your 42hrs seem reasonable for this type of cpu. Will give you some better runtime estimates tomorrow after I got more of this WU done. |
Send message Joined: 19 Jul 10 Posts: 589 Credit: 18,926,809 RAC: 4,448 |
Below the SSE2 path there is only x87, so there is nothing going wrong. I think cruncher said there wasn't really a difference between SSE and X87 with the functions he was giving to Matt. Thanks for all this information, this part was the most important for me since I wasn't sure, if the application runs as it should. The GPU app (or BOINC) on my other machine didn't recognize SSE3 when it was still on WinXP, so I thought here might also be something wrong. |
Send message Joined: 19 Jul 10 Posts: 589 Credit: 18,926,809 RAC: 4,448 |
That was looong WU... Looks like the stock app could use some optimizing, the old Gipsel app did a WU with similar amount of credit in about 1/4 of time. |
©2024 Astroinformatics Group