|
|
Ublas : |
Subject: Re: [ublas] [PATCH 3/3] boost::ublas increasing the range of BLAS level 3 benchmarks
From: Nasos Iliopoulos (nasos_i_at_[hidden])
Date: 2016-03-11 08:20:15
Regardless, these are great figures.
Can you please run them comparing the simple uBlas implementation for
matrices from 2 to 100 with the gemm based one with a signle thread? I
wonder when the control statement starts to play a role.
What do you think should be the plan to switch from multi-core to to
single-threaded so as to not get all the communication hit for smaller
matrices?
- Nasos
On 03/10/2016 03:14 PM, palik imre wrote:
> Forget OpenMP for the time being.
>
> For this gemm implementation to scale beyond a socket boundary, one
> needs OpenMP version 4.0 or better, and some way to determine the
> processor topology.
>
> I have no idea how to determine the processor topology in a portable
> way (or any way outside Linux). If somebody is willing to help with
> this, then we can pull it off. Otherwise I cannot do it.
>
> BTW, some measurements of the basic idea on a 2*8 core Ivy Bridge:
>
> Single threaded:
> # m uBLAS: t1 MFLOPS Blocked: t2 MFLOPS Diff nrm1
> 100 0.000231629 8634.5 0.000179467 11144.1 0 0.000139046
> 14383.7 0
> 200 0.00146996 10884.6 0.000969961 16495.5 0 0.000941657
> 16991.3 0
> 300 0.00366689 14726.4 0.00328796 16423.6 0 0.00327107
> 16508.4 0
> 400 0.00909807 14068.9 0.00871394 14689.1 0 0.0087016
> 14709.9 0
> 500 0.017248 14494.4 0.0167666 14910.6 0 0.0167691
> 14908.3 0
> 600 0.0289742 14909.8 0.0282574 15288 0 0.0282521
> 15290.9 0
> 700 0.0462171 14843 0.0445143 15410.8 0 0.044418
> 15444.2 0
> 800 0.0678457 15093.1 0.0645271 15869.3 0 0.065332
> 15673.8 0
> 900 0.0998233 14605.8 0.0974452 14962.3 0 0.0977101
> 14921.7 0
> 1000 0.133766 14951.4 0.131982 15153.6 0 0.132004
> 15151.1 0
> 1100 0.177154 15026.5 0.175536 15165 0 0.175679
> 15152.7 0
> 1200 0.226595 15251.9 0.223401 15469.9 0 0.223272
> 15478.9 0
> 1300 0.281128 15629.9 0.279018 15748.1 0 0.279577
> 15716.6 0
> 1400 0.358121 15324.4 0.358763 15297 0 0.356956
> 15374.5 0
> 1500 0.441216 15298.6 0.44028 15331.2 0 0.440745
> 15315 0
> 1600 0.539017 15198 0.540438 15158.1 0 0.538543
> 15211.4 0
> 1700 0.631606 15557.2 0.63106 15570.6 0 0.632818
> 15527.4 0
> 1800 0.740764 15745.9 0.740261 15756.6 0 0.740176
> 15758.4 0
> 1900 0.885414 15493.3 0.885684 15488.6 0 0.885026
> 15500.1 0
> 2000 1.04361 15331.3 1.04518 15308.4 0 1.04733
> 15277 0
>
> 8 threads on one socket:
> # m uBLAS: t1 MFLOPS Blocked: t2 MFLOPS Diff nrm1
> 100 0.0013259 1508.4 0.000470025 4255.09 0 0.0004592
> 4355.4 0
> 200 0.000638351 25064.6 0.000245947 65054.7 0 0.000244369
> 65474.8 0
> 300 0.000895587 60295.6 0.000792129 68170.7 0 0.000760058
> 71047.2 0
> 400 0.00231167 55371.3 0.00189499 67546.4 0 0.00190057
> 67348.1 0
> 500 0.003532 70781.4 0.00312475 80006.4 0 0.00313642
> 79708.8 0
> 600 0.00603096 71630.4 0.00529135 81642.7 0 0.00535545
> 80665.5 0
> 700 0.00961982 71311.1 0.00780231 87922.7 0 0.00788078
> 87047.2 0
> 800 0.0143081 71567.8 0.0119397 85764.6 0 0.0123294
> 83053.5 0
> 900 0.0195524 74568.7 0.0175708 82978.6 0 0.0178417
> 81718.6 0
> 1000 0.0245773 81376 0.023048 86775.3 0 0.0233121
> 85792.3 0
> 1100 0.0321172 82883.9 0.0305309 87190.5 0 0.0307586
> 86544.8 0
> 1200 0.0388261 89012.2 0.0372826 92697.5 0 0.0376794
> 91721.1 0
> 1300 0.0479156 91702.8 0.046001 95519.7 0 0.0464177
> 94662.1 0
> 1400 0.0590871 92879.8 0.0571302 96061.3 0 0.0597637
> 91828.3 0
> 1500 0.0717213 94114.3 0.0697929 96714.7 0 0.0699795
> 96456.8 0
> 1600 0.0844559 96997.4 0.084161 97337.3 0 0.0837095
> 97862.3 0
> 1700 0.0980906 100173 0.0979744 100291 0 0.0979657
> 100300 0
> 1800 0.116091 100473 0.115783 100740 0 0.115882
> 100654 0
> 1900 0.135347 101354 0.134988 101624 0 0.136053
> 100828 0
> 2000 0.162126 98688.5 0.161714 98940.2 0
> 0.162389 98529.1 0
>
> 16 threads on 2 sockets:
> # m uBLAS: t1 MFLOPS Blocked: t2 MFLOPS Diff nrm1
> 100 0.00402575 496.801 0.00264263 756.823 0
> 0.00271382 736.969 0
> 200 0.00266775 5997.56 0.00222018 7206.64 0
> 0.00172627 9268.53 0
> 300 0.0037985 14216.1 0.00345001 15652.1 0
> 0.00337945 15978.9 0
> 400 0.00516891 24763.4 0.00502001 25498 0
> 0.00446124 28691.6 0
> 500 0.00761118 32846.4 0.00615818 40596.4 0
> 0.00594556 42048.2 0
> 600 0.0116084 37214.5 0.00914148 47257.1 0
> 0.0091653 47134.3 0
> 700 0.012626 54332.5 0.00967706 70889.3 0
> 0.0105016 65323.2 0
> 800 0.0161321 63476.1 0.01418 72214.5 0
> 0.0138772 73790.1 0
> 900 0.0169242 86148.9 0.0155385 93831.5 0
> 0.0152458 95632.8 0
> 1000 0.0208066 96123.5 0.0171705 116479 0
> 0.0178461 112069 0
> 1100 0.0399168 66688.7 0.0388316 68552.4 0
> 0.0383423 69427.2 0
> 1200 0.0448104 77124.9 0.0422857 81729.8 0
> 0.0425542 81214.1 0
> 1300 0.0509146 86301.4 0.0469193 93650.2 0
> 0.047686 92144.5 0
> 1400 0.0558425 98276.3 0.0579064 94773.7 0
> 0.0516464 106261 0
> 1500 0.0634977 106303 0.0638868 105656 0
> 0.0598486 112785 0
> 1600 0.0703613 116428 0.0725247 112955 0
> 0.0683293 119890 0
> 1700 0.0803688 122261 0.0790288 124334 0
> 0.07428 132283 0
> 1800 0.0860819 135499 0.089678 130065 0
> 0.0819196 142383 0
> 1900 0.098444 139348 0.0951359 144194 0
> 0.090089 152272 0
> 2000 0.107054 149457 0.120315 132985 0
> 0.102137 156652 0
>
> all three versions are gemm-based.
>
> On Wednesday, 9 March 2016, 14:43, Nasos Iliopoulos
> <nasos_i_at_[hidden]> wrote:
>
>
> I think the matrix abstraction the way it is now is ok. It would be
> confusing to have a large matrix class.
>
> My current thinking is that we should have a two tier switch. One that
> detects that openmp is enabled and one that enables parallelization
> based on user preference:
>
> #ifdef _OPENMP && BOOST_UBLAS_PARALLEL
> // parallel code (with runtime switch if needed) goes here
> #else
> #ifdef BOOST_UBLAS_PARALLEL
> #warning "OPENMP not present. boost::ublas parallel mode not enabled."
> #end
> // serial code goes here
> #endif
>
> to enable parallel mode:
> gcc myfile.cpp -o myexe -fopenmp -DBOOST_UBLAS_PARALLEL
>
> the following does not enabe ublas parallel mode but let the user's
> openmp code run:
> gcc myfile.cpp -o myexe -fopenmp
>
> this will not enable parallelization at all:
> gcc myfile.cpp -o myexe
>
>
> essentially _OPENMP is defined when you pass the -fopenmp argument to
> gcc and I suppose in all other compilers that support the standard.
>
> * One downside of this approach is that temporarily disabling ublas
> parallel mode would need some hoci poci.
>
> * I think that this approach is better than nothing and If you can
> think of a more clear and/or efficient way please voice it.
>
> * I would favor the std::thread approach but thinking about it again I
> believe we will need to introduce state so the we have a facility to
> define the number of threads.We could use
> (http://en.cppreference.com/w/cpp/utility/program/getenv) but this
> wouldn't allow for after-execution changes. On the other hand openmp
> has state and the user can use it deliberately.
>
> -Nasos
>
>
>
>
> _______________________________________________
> ublas mailing list
> ublas_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
> Sent to: nasos_i_at_[hidden]