Comparison of clients

Forums › Cruncher's Corner

StoneLord
StoneLord
Joined: 15 Jun 05
Posts: 13
Credit: 50798
RAC: 0
5 Apr 2006 14:46:26 UTC
Topic 191034
(moderation:
    )

    This is the small comparison test of akosf optimized clients-> albert_4.37_x86 for Windoze

    CPU,RAM,Instruction set:

    AMD Athlon XP (Barton) 3200+@2200 MHz
    512MB RAM
    MMX(+),3DNow!(+),SSE

    Tested WU:
    (short) z1_0261.5_2539_S4R2a_1

    -------------------------------------------------
    |version |time[s]| curr./orig.| speedup |
    | | | | |
    |-----------------------------------------------|
    |original | 4553 | 1.0000 | 1.0000 |
    |-----------------------------------------------|
    |387 | 2294 | 0.5038 | 1.9847 |
    |C-37 | 1867 | 0.4100 | 2.4387 |
    |C-40 | 1653 | 0.3630 | 2.7544 |
    |-----------------------------------------------|
    |A-36 | 1944 | 0.4269 | 2.3421 |
    |D-40 | 1194 | 0.2622 | 3.8132 |
    |-----------------------------------------------|
    |S-38 | 1354 | 0.2974 | 3.3626 |
    |S-39 | 1389 | 0.3051 | 3.2779 |
    |S-39L | 1249 | 0.2743 | 3.6453 |
    |S-40 | 1144 | 0.2512 | 3.9799 |
    -------------------------------------------------

    So the best results gives S-40 with SSE 3.9799 x faster than original
    second place D-40 with 3DNow! 3.8732 x faster than original

    Thanks akosf, you have done great job !!!!

    archae86
    archae86
    Joined: 6 Dec 05
    Posts: 3157
    Credit: 7205754931
    RAC: 939668

    Comparison of clients

    5 Apr 2006 16:05:00 UTC
    Message 27405
    (moderation:
      )

      Quote:
      This is the small comparison test of akosf optimized clients-
      Tested WU:
      (short) z1_0261.5_2539_S4R2a_1

      So you actually did a controlled comparison of the science aps (not clients) using a single standard work unit. Outstanding.

      Thanks for the data. My guess is that your method gives a more accurate comparison for most purposes than those of us providing before/after results on whatever work units we happen to get.

      If there is meaningful variation in computation type content among WU's, you fail to randomize that, but you should get a far more accurate comparision of the actual work at hand. I suspect Einstein does not vary much in computation mix (floating vs integer mix, locality of reference ...) from WU to WU, though I have my doubts on that score with SETI.

      Would you care to share your method for "capturing" a work unit and running and measuring the time "privately". I assume you disengage the boincmgr and boinc.exe, but don't know just how you handle this. Perhaps some other folks here might try your method on some of the other CPU types.

      Akos Fekete
      Akos Fekete
      Joined: 13 Nov 05
      Posts: 561
      Credit: 4527270
      RAC: 0

      RE: Thanks akosf, you have

      5 Apr 2006 16:33:07 UTC
      Message 27406
      (moderation:
        )

        Quote:
        Thanks akosf, you have done great job !!!!

        Thanks.
        I did a modification on S40, I hope it will be reach the 4.000 ratio... :-)

        Steve Cressman
        Steve Cressman
        Joined: 9 Feb 05
        Posts: 104
        Credit: 139654
        RAC: 0

        RE: RE: Thanks akosf, you

        5 Apr 2006 17:03:48 UTC
        Message 27407 in response to message 27406
        (moderation:
          )

          Quote:
          Quote:
          Thanks akosf, you have done great job !!!!
          Thanks.
          I did a modification on S40, I hope it will be reach the 4.000 ratio... :-)


          Are you saying there are two versions of S40 now? And those that got it b4 should d/l again???
          :)

          98SE XP2500+ @ 2.1 GHz Boinc v5.8.8

          Akos Fekete
          Akos Fekete
          Joined: 13 Nov 05
          Posts: 561
          Credit: 4527270
          RAC: 0

          RE: RE: I did a

          5 Apr 2006 17:47:59 UTC
          Message 27408 in response to message 27407
          (moderation:
            )

            Quote:
            Quote:
            I did a modification on S40, I hope it will be reach the 4.000 ratio... :-)
            Are you saying there are two versions of S40 now? And those that got it b4 should d/l again???
            :)

            Pooh! I didn't change the version because this modification is very small.
            But I will change it before upload. Okay?
            What do you think about a fine codename? :-)

            Steve Cressman
            Steve Cressman
            Joined: 9 Feb 05
            Posts: 104
            Credit: 139654
            RAC: 0

            RE: RE: RE: I did a

            5 Apr 2006 18:01:20 UTC
            Message 27409 in response to message 27408
            (moderation:
              )

              Quote:
              Quote:
              Quote:
              I did a modification on S40, I hope it will be reach the 4.000 ratio... :-)
              Are you saying there are two versions of S40 now? And those that got it b4 should d/l again???
              :)
              Pooh! I didn't change the version because this modification is very small.
              But I will change it before upload. Okay?
              What do you think about a fine codename? :-)


              How about S40.01 or S40again or S40.4x
              Yes, S40.4x. The 4x for 4 times faster, lol.
              :)

              98SE XP2500+ @ 2.1 GHz Boinc v5.8.8

              Akos Fekete
              Akos Fekete
              Joined: 13 Nov 05
              Posts: 561
              Credit: 4527270
              RAC: 0

              RE: RE: What do you think

              5 Apr 2006 18:20:31 UTC
              Message 27410 in response to message 27409
              (moderation:
                )

                Quote:
                Quote:
                What do you think about a fine codename? :-)

                How about S40.01 or S40again or S40.4x
                Yes, S40.4x. The 4x for 4 times faster, lol.
                :)

                S40.01 sounds well, because I have new ideas. (small things, but lot)
                And two decimal places will be enough. (we will run out of work in some days!)

                StoneLord
                StoneLord
                Joined: 15 Jun 05
                Posts: 13
                Credit: 50798
                RAC: 0

                RE: RE: This is the small

                5 Apr 2006 20:38:52 UTC
                Message 27411 in response to message 27405
                (moderation:
                  )

                  Quote:
                  Quote:
                  This is the small comparison test of akosf optimized clients-
                  Tested WU:
                  (short) z1_0261.5_2539_S4R2a_1

                  So you actually did a controlled comparison of the science aps (not clients) using a single standard work unit. Outstanding.

                  Thanks for the data. My guess is that your method gives a more accurate comparison for most purposes than those of us providing before/after results on whatever work units we happen to get.

                  If there is meaningful variation in computation type content among WU's, you fail to randomize that, but you should get a far more accurate comparision of the actual work at hand. I suspect Einstein does not vary much in computation mix (floating vs integer mix, locality of reference ...) from WU to WU, though I have my doubts on that score with SETI.

                  Would you care to share your method for "capturing" a work unit and running and measuring the time "privately". I assume you disengage the boincmgr and boinc.exe, but don't know just how you handle this. Perhaps some other folks here might try your method on some of the other CPU types.

                  I used BOINC 5.2.13 optimized by truX

                  I download some WUs, then i stopped crunching any unit, suspend all unit except z1_0261.5_2539_S4R2a_1 and all network activity, then I create some copies of BOINC directory for every tested "science aps", then each of tested "albert_4.37_windows_intelx86.exe" original and akosfs versions were copied to einstein directories, so it was 10 directories.

                  After that, i started one of my copied BOINCmgr and also start only that one WU which was not suspended. When WU was cruched i wrote the time into tablet. So i did this 10 times.

                  It is quite simple method, but i think the results are usable.

                  Thats all...

                  archae86
                  archae86
                  Joined: 6 Dec 05
                  Posts: 3157
                  Credit: 7205754931
                  RAC: 939668

                  RE: I download some WUs,

                  5 Apr 2006 21:17:42 UTC
                  Message 27412 in response to message 27411
                  (moderation:
                    )

                    Quote:

                    I download some WUs, then i stopped crunching any unit, suspend all unit except z1_0261.5_2539_S4R2a_1 and all network activity, then I create some copies of BOINC directory for every tested "science aps", then each of tested "albert_4.37_windows_intelx86.exe" original and akosfs versions were copied to einstein directories, so it was 10 directories.

                    After that, i started one of my copied BOINCmgr and also start only that one WU which was not suspended. When WU was cruched i wrote the time into tablet. So i did this 10 times.

                    It is quite simple method, but i think the results are usable.

                    Thats all...

                    Got it. The suspended network activity would stop spurious reporting of multiple results for the same WU at the time. For cleanup in the end is just deleting the replicated BOINC directories enough? Or is there something more global than that to assure a clean restart when you resume real processing?

                    StoneLord
                    StoneLord
                    Joined: 15 Jun 05
                    Posts: 13
                    Credit: 50798
                    RAC: 0

                    RE: RE: Got it. The

                    5 Apr 2006 21:49:34 UTC
                    Message 27413 in response to message 27412
                    (moderation:
                      )

                      Quote:
                      Quote:

                      Got it. The suspended network activity would stop spurious reporting of multiple results for the same WU at the time. For cleanup in the end is just deleting the replicated BOINC directories enough? Or is there something more global than that to assure a clean restart when you resume real processing?

                      Yes, It is enough. After testing, all replicated BOINC were deleteted. (but i still have one for new akosf optimized app.).
                      For example, I did not test (S-37a) because of SSE2.

                      Georgina
                      Georgina
                      Joined: 16 Feb 05
                      Posts: 5
                      Credit: 352151
                      RAC: 0

                      I don't understand the

                      5 Apr 2006 23:12:12 UTC
                      Message 27414 in response to message 27413
                      (moderation:
                        )

                        I don't understand the naming. D40 vs S40, etc.

                        I am running D40 on my AMD XP PCs. This is correct? or should I switch to the S40 for a little more speed?

                        G

                        Comment viewing options

                        Select your preferred way to display the comments and click "Save settings" to activate your changes.

                        Forums › Cruncher's Corner