Message boards :
Number crunching :
Windows port of Alex v8 code
Message board moderation
Previous · 1 . . . 37 · 38 · 39 · 40 · 41 · 42 · 43 . . . 50 · Next
Author | Message |
---|---|
David Send message Joined: 19 May 99 Posts: 411 Credit: 1,426,457 RAC: 0 |
I will pull the plug on Ginger when her cache empties in a couple days. Jeez, sounds rather dramatic. What did Ginger do to you to deserve such a nasty fate... :) My focus is now pointed at "burning in" the release candidates so there are fewer (hopefully none), surprises upon release. Only real way is to widen the test group. Look what happens with each Microsoft release as its tested with a small and knowledgeable userbase, but give it to the idiots out in public and they break it quicker than a McDonalds Toy. and have refrained from cancelling WUs where my wingam was running BOINC 4.45. I dont look that hard, but have been caught by a few running the old version. Mt RAC is jumping by a few hundred now & then, but I blame the WU's rather than anything else. Plus I had a heap on the David_Quad_core that failed for some reason, but took a few thousand seconds to do it, and gave me 0 credits - argh |
JDWhale Send message Joined: 6 Apr 99 Posts: 921 Credit: 21,935,817 RAC: 3 |
I will pull the plug on Ginger when her cache empties in a couple days. Summer is approaching quickly here in SE Texas. Room temps are at 27C and Ginger stays in heat. Coupled with her energy draw puts me over my project budget of 1kW since Wrongway came online. She may be going down, but she is not going out. Is destined to my 21YO daughter come June. My focus is now pointed at "burning in" the release candidates so there are fewer (hopefully none), surprises upon release. I'm trying everything I know to break RC1... Used it as a coffee holder one morning... Had it fetch the post another...Manning the BBQ last night... Solid as a rock! We're talking "Top Shelf" stuff here. and have refrained from cancelling WUs where my wingam was running BOINC 4.45. I hate those... and sometimes you get a gaggle of WUs with the same wingman, who happens to go missing leaving you hold the pending bag for a couple months. Yes, modus operandi(MO) for Thurston avoid most "pending" pitfalls as credit is usually granted when the result is returned ("checked but no concensus" is the exception). As such I would expect RAC to stabilize 2 weeks after change is applied as long as he doesn't run out of work. That is one of the pitfalls of this MO. Without many pending, his RAC will drop like a rock when the work runs out... Thus a large cache is essential. Without another interruption, I expect Thurston to reach stable RAC in about 72-96 hours, target is still RAC 7090. G'Day, JDWhale |
mr.kjellen Send message Joined: 4 Jan 01 Posts: 195 Credit: 71,324,196 RAC: 0 |
I see a Whale roaming in the waters of the top 20...Congrats JD. /Anton |
Francois Piednoel Send message Joined: 14 Jun 00 Posts: 898 Credit: 5,969,361 RAC: 0 |
Starting this morning, one of my Skulltrail will be running John (JDWhale) version of the Alex V8 code, based on his request. I will use the other machine to try to catch up with my own code :) , that is a big task, because Alex/John code is amazingly fast. I am so busy preparing the next step of the hardware ... We will be able to see how my actual design can do on the best code ;) Francois |
JDWhale Send message Joined: 6 Apr 99 Posts: 921 Credit: 21,935,817 RAC: 3 |
Starting this morning, one of my Skulltrail will be running John (JDWhale) version of the Alex V8 code, based on his request. I will use the other machine to try to catch up with my own code :) , that is a big task, because Alex/John code is amazingly fast. Thank you for allowing AK_WhalePort to run on your Skulltrail. The version I provided is V0.2X built using SSE4.1 IPP 32-bit. Since I do not have any hardware to test this version (45nm chip) I can only hope that this client version will perform well. I will be watching and waiting to see some results post. Note: AK_WhalePort is a "verbatim" windows port of Alex v-8 code... I only did what was necessary to get the source code to build & link substituting Intel IPP calls for Apple proprietary DSP functions. Again... Thank you Francois.. Best regards, JDWhale @mr.kjellen - Thanks... except for untold misfortune by the leaders I can not expect Thurston to rise any more... Since starting this quest/race the RAC to reach top 20 has risen by maybe 300-500. |
Francois Piednoel Send message Joined: 14 Jun 00 Posts: 898 Credit: 5,969,361 RAC: 0 |
Starting this morning, one of my Skulltrail will be running John (JDWhale) version of the Alex V8 code, based on his request. I will use the other machine to try to catch up with my own code :) , that is a big task, because Alex/John code is amazingly fast. It is now running, not generating errors, the L2 cache miss are much lower than before, nice work from Alex and team. this is the machine you want to look at: hostid=4081480 I am jalous :) V / Who? / Francois |
W-K 666 Send message Joined: 18 May 99 Posts: 19078 Credit: 40,757,560 RAC: 67 |
Right, I'm setup and waiting, patiently, for these new apps. Just connected new host Q9450 at 15:42 UTC first two results returned. Still at stock speed etc. Actually my sons, but haven't told him components and new 24" widescreen monitor are here yet, but he has paid me, so maybe I better. It will be connected via his account when he takes it to his place. |
John Clark Send message Joined: 29 Sep 99 Posts: 16515 Credit: 4,418,829 RAC: 0 |
Starting this morning, one of my Skulltrail will be running John (JDWhale) version of the Alex V8 code, based on his request. I will use the other machine to try to catch up with my own code :) , that is a big task, because Alex/John code is amazingly fast. Thank you for allowing AK_WhalePort to run on your Skulltrail. The version I provided is V0.2X built using SSE4.1 IPP 32-bit. Since I do not have any hardware to test this version (45nm chip) I can only hope that this client version will perform well. I will be watching and waiting to see some results post. Excellent addition to see how the Alex code port, care of JD, will perform. @mr.kjellen - Thanks... except for untold misfortune by the leaders I can not expect Thurston to rise any more... Since starting this quest/race the RAC to reach top 20 has risen by maybe 300-500. It is interesting to see how the top 20 entry requirement has changed in the last 12 months. My first Quad was a QX6,700 clocked to 3.1GHz from stock. This stabilised at it's current RAC while reaching a maximum host position of 19 in the top 20. So, it looks like the entry RAC is heading to double that required 12 - 14 months ago. It's good to be back amongst friends and colleagues |
ML1 Send message Joined: 25 Nov 01 Posts: 20331 Credit: 7,508,002 RAC: 20 |
Starting this morning, one of my Skulltrail will be running John (JDWhale) version of the Alex V8 code, based on his request. I will use the other machine to try to catch up with my own code :) ... A most excellent test... But why only 32-bit? Good luck and may the best code win! (Please, just the results ;-) ) Happy crunchin, Martin See new freedom: Mageia Linux Take a look for yourself: Linux Format The Future is what We all make IT (GPLv3) |
JDWhale Send message Joined: 6 Apr 99 Posts: 921 Credit: 21,935,817 RAC: 3 |
Starting this morning, one of my Skulltrail will be running John (JDWhale) version of the Alex V8 code, based on his request. I will use the other machine to try to catch up with my own code :) ... First results have posted hostid=4081480&offset=1620 These early results are hybrids, partially crunched with another client. "Purebreed" should be a bit faster ;-) @Martin - I only DL and install Intel ICC & IPP 32-bit evaluation copies. So the "real" clients will be even faster. In addition, 32-bit runs on both 32 & 64 bit OS. |
Geek@Play Send message Joined: 31 Jul 01 Posts: 2467 Credit: 86,146,931 RAC: 0 |
With the release of the new app's about 1 week away, I have one question. Can someone explain in layman's terms........Why is Alex Kan's V8 source code so much faster than what we had before? |
Francois Piednoel Send message Joined: 14 Jun 00 Posts: 898 Credit: 5,969,361 RAC: 0 |
|
Francois Piednoel Send message Joined: 14 Jun 00 Posts: 898 Credit: 5,969,361 RAC: 0 |
With the release of the new app's about 1 week away, I have one question. From what I can tell with vtune, Alex code did improve dramatically the data locality, The L2 cache miss ratio decrease by a factor of 10, and L1 cache is much more efficent.Alex can probably confirm it, I don t have the "accelerator.h", so, I can t confirm from a code point of view. Who? |
JDWhale Send message Joined: 6 Apr 99 Posts: 921 Credit: 21,935,817 RAC: 3 |
With the release of the new app's about 1 week away, I have one question. I'll try one example... When wanting to find the MEAN of a vector(1D array), typically a programmer would write ~5 lines of code to loop over the all the elements to gather a sum, then divide the sum by the vector length. Simple enough and easy to read/understand, even for most non-programmers. Alex code might implement the same operation using maybe 100 lines of code, split the array into 4 or eight pipes, feed the "vector" part of the processor allowing 4 or 8 sums to take place in the same amount of CPU clicks as a single sum operation. Accumulate the sum from the separate pipes, then do the division. Not very intuitive, but making much more efficient use of the CPU and making the code much more complex to read and to write and to comprehend. Also requires coding some of the optimizations multiple times, once for each CPUs capabilities to gain maximum performance on each CPU type. This process uses what are called "intrinsic functions" that allow "low level" access to the CPU capabilities. This set of intrinsics change from CPU to CPU and are sometimes refered to as SSE, SSE2, SSE3, SSSE3, SSE4.1, MMX, ...etc. Explaining the need for more than one client, depending on which "instructions" your host(CPU) can understand. This was just the most simple of examples of how Alex speed up the SETI client. Regards, JDWhale [edit] The Skulltrail has changed hostid and "detached"... hostid=4262769&offset=1800 [/edit] |
JDWhale Send message Joined: 6 Apr 99 Posts: 921 Credit: 21,935,817 RAC: 3 |
[edit]message deleted[/edit] The Skulltrail has changed hostid and "detached"... New host is hostid=4262769 |
Sir Ulli Send message Joined: 21 Oct 99 Posts: 2246 Credit: 6,136,250 RAC: 0 |
|
Fred W Send message Joined: 13 Jun 99 Posts: 2524 Credit: 11,954,210 RAC: 0 |
|
Francois Piednoel Send message Joined: 14 Jun 00 Posts: 898 Credit: 5,969,361 RAC: 0 |
[edit]message deleted[/edit] that was weird ... Can't explain what happenned... Francois |
JDWhale Send message Joined: 6 Apr 99 Posts: 921 Credit: 21,935,817 RAC: 3 |
[edit]message deleted[/edit] If you crashed... Mark had mentioned he needed to raise vCore a bit to account for the extra stress being put on his OC penryn with the new client. You probably have more information/resources on that requirement than I could help with. Check your BOINC log. About 4 days ago Thurston suffered from a "random project reset" where all WUs were deleted and detatched. More details on this were posted to this thread... like I said about 4 days ago. Since then I pointed a small fan directly on my memory and adjusted vDimm18 a bit. I don't have a lot of confidence in my BALLISTIX DDR2-1066 dimms. Looking good, John |
David Send message Joined: 19 May 99 Posts: 411 Credit: 1,426,457 RAC: 0 |
Since then I pointed a small fan directly on my memory and adjusted vDimm18 a bit. I don't have a lot of confidence in my BALLISTIX DDR2-1066 dimms. I recently bought 2 1GB sticks as well and it took less than 2 days before I removed them and went back to the Geil 800 sticks - even running under 1066 the Ballistix showed errors and poor memory speed with various memory tests and benchmark programs. Hard to imagine why 1066 ram is slower than 800 (at 840) ram. Ah well, such is life. |
©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.