1) Message boards : Webpage and Application Code Discussion : EXACT on GitHub (Message 7624)
Posted 12 Oct 2018 by Tetsuji Maverick Rai
Hi,

I tried to compile in many ways with Intel C/C++ compiler on Fedora 28 with AVX2 on, but can't beat the stock program....Is it built with AMD optimizing compiler? But Intel compiler should be faster... I wonder why. I tried ubuntu 16.04 and 18.04 also (installed them in vmware), but to no avail.

Will you tell me any suggestions? If I can build one, I will make it public.

Thanks!

EDIT: This git repository seems to be modified often.... So I went back to Sep 16 2017 on git, but today's git won't compile even after that! I suspect my problem might be due to modified git....
2) Message boards : News : [all] status update (Message 7612)
Posted 26 Sep 2018 by Tetsuji Maverick Rai
Hi Everyone!

Sorry for a lot of silence on our end -- things have been quite hectic this summer! I accepted a new position and now am at Rochester Institute of Technology and busily getting a new lab set up (and having some students migrate along with me).

I'm going to let some of the work units clear out of the system and set up some new runs which should help fix some of the issues you've been seeing with bad workunits. In the longer term, I've started work on an updated binary which should also allow for the use of GPUs, which should hopefully allow the runs to scale a bit better without running into these problems.

I'll be keeping you posted as that comes along, but right now we're all just trying to get settled in a new place! Thanks again for all your support. The results coming in are definitely interesting ones!


Thank you for update Travis, I wish you a good luck at your new lab!!

One suggestion. Banning bad hosts is one way, but I want to increase "max # of total tasks" for each workunit to about 10. Then however many bad results come, a correct result will come before 10 and the workunit will be validated later instead of being trashed. Other projects like Amicable numbers do it and they can handle with error results without banning. Fortunately this server seems to have a good internet bandwidth, so no ban is necessary.

Just my thought.

Best regards,

-Tetsuji
3) Message boards : Webpage and Application Code Discussion : EXACT on GitHub (Message 7529)
Posted 23 May 2018 by Tetsuji Maverick Rai
I found an interesting machine. In workunit https://csgrid.org/csg/workunit.php?wuid=3729548, which was crunched by two i7 7700 machines, elapsed times are quite different. The difference is OS. Upper faster one may be overclocked. Very unusual. Sorry, this post is also my memo.
4) Message boards : Webpage and Application Code Discussion : EXACT on GitHub (Message 7527)
Posted 23 May 2018 by Tetsuji Maverick Rai
Now I'm testing it. At first it gave several invalid results, but the latest one (older source from git and not so aggressive optimization) looks safe, but it's not confirmed yet.

When it gives 100 valid results in a row (currently 40+), if someone needs one, I will upload it, maybe in my google drive along with app_info.xml. But only for Haswell or up (it means it must support AVX2) for Linux. Not for Windows because I don't have Intel compiler for Windows. On AMD processors it will work fine, but Intel takes full advantage of it.

If anyone wants, I'll make SSE3 version also. I guess SSE3 version and AVX2 version have little difference in speed. The compiler itself produces up to AVX512, but since my machine supports up to AVX2, I cannot compile AVX512. BOINC demands to run the produced binary on the product machine upon configuration. If you want "incomplete" AVX512 version, I can make one. CSGRID doesn't require me to test run the binary it produces.

BTW I don't like benchmark test. Someone please do it if necessary. So far, it looks like mine (4790K at 4.0GHz non-turbo) is faster than i7 8700K (turbo or not unknown) for each process. But somehow mine was slower than i7 7700. See https://csgrid.org/csg/workunit.php?wuid=3730109. Upper is mine, lower is i7 7700.
5) Message boards : Webpage and Application Code Discussion : EXACT on GitHub (Message 7520)
Posted 19 May 2018 by Tetsuji Maverick Rai
Hi Travis,

Thank you for your hard work.

I am new to this project. I tried to make an optimized client with that source on Fedora 27 Linux with Intel Core i7-4790K.

At first I need to set environment variable CC,CXX,BOINC_SOURCE so that I can use Intel compiler and boinc source, and edited exact/CMakeLists.txt so that a client was compiled and set compile flags properly.

Then, on exact/build build by me, ran "cmake .." and "make exact_client_0.33".

I created app_info.xml, and ran that client. I found the name of application was "exact_bn_sfmp". Strangely, if the version number begins with zero, it was regarded as an octal number, so I needed to write 041 to set it to 0.33.

Somehow, estimated times of most tasks are longer than 1 day or some are longer than 4 days, while they are usually 1.5 to some hours with stock application. Calculation seems to be correct, but why is the estimated time so long? I confirmed compiler flags were working correctly with "ps ax".

Thanks in advance!!

EDIT: Anyway, I will try to test it. Initial estimated times are often unreliable.

EDIT2: Actually they are proceeding steadily, maybe faster than normal. Thank you very much!