You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@hama.apache.org by Victor Lee <vv...@gmail.com> on 2014/09/23 06:16:38 UTC
PageRank Benchmark Test (0.6.4 vs 0.7.0-SNAPSHOT)
Hello all.
I conducted a perfromance test about latest version 0.7.0 of Apache Hama
(applied Kryo)
(Kryo is a fast and efficient object graph serialization framework for Java)
The results of the test, the performance has been greatly improved(20~25%).
Kryo's serialization API is speed, efficiency and an easy,
so useful for hama's network communication.
It seems a very good improvement!. : )
I will attach the test results below.
Thanks.
[image: 埋め込み画像 1]
( supported by Byungseok Min. Thanks ! )
Re: PageRank Benchmark Test (0.6.4 vs 0.7.0-SNAPSHOT)
Posted by Victor Lee <vi...@apache.org>.
Hi. Edward.
Thanks for explaining about detail improvement.
If 'HAMA-919' is done, I will re-test. ; )
2014-09-24 18:16 GMT+09:00 Edward J. Yoon <ed...@apache.org>:
> Thanks for your benchmarks. BTW, I guess I need to explain more about this.
>
> As I described HAMA-919, we have to manage the outgoing messages per
> vertex (like giraph), and won't use sorted queue for graph job
> anymore. This is the way to improve the performance and memory
> efficiency. With this, I think VerticesInfo also can be changed. So
> now, it's a meaningless testing. I kindly ask you to benchmark again
> after HAMA-919 is done.
>
> However, one thing is clear that Kryo serializer is really useful. :-)
>
> On Tue, Sep 23, 2014 at 2:32 PM, Victor Lee <vv...@gmail.com> wrote:
> > Hello, All ~
> >
> > I didn't send attach file. so I send mail again.
> >
> > thanks.
> >
> >
> >
> >
> > [ Apache Hama Benchmark Test ]
> > . PageRank Benchmark Test 0.6.4 vs 0.7.0-SNAPSHOT (2014/09/20)
> > . Environments
> > - 12 core (2.66Ghz 6core * 2)
> > - 32G Memory (4G * 8)
> > - 2 Node Hama Cluster
> > - 6 tasks (opt : -Xmx1024m, -Xms2048)
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > . Memory : -Xmx1024m
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > vertices Edges
> > per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
> > rate(%)
> > 1th 2nd 3rd average 1th 2nd 3rd Average
> > (seconds)
> > 40000 500 178.822 178.688 183.85 180.453 147.711 142.232 132.755 140.899
> > -21.92
> > 40000 600 231.884 231.964 226.112 229.987 177.764 174.733 171.714 174.737
> > -24.02
> > 40000 700 277.061 276.845 273.928 275.945 219.796 222.792 216.817 219.802
> > -20.35
> > 40000 800 Job failed. Job failed.
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > . Memory : -Xmx2048m
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > vertices Edges
> > per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
> > rate(%)
> > 1th 2nd 3rd average 1th 2nd 3rd Average
> > (seconds)
> > 40000 500 168.741 171.706 168.701 169.716 132.737 130.850 126.670 130.086
> > -23.35
> > 40000 600 207.749 208.045 195.854 203.883 159.744 153.757 156.681 156.727
> > -23.13
> > 40000 700 240.848 240.837 232.015 237.900 181.217 186.806 187.221 185.081
> > -22.20
> > 40000 800 280.729 274.51 280.754 278.664 210.918 209.112 217.013 212.348
> > -23.80
> > 40000 900 320.428 320.683 314.3 318.470 235.184 242.585 238.244 238.671
> > -25.06
> > 40000 1000 355.425 361.342 358.36 358.376 276.705 279.91 281.294 279.303
> > -22.06
> > 40000 1100 394.008 401.419 399.351 398.259 307.003 304.691 306.556
> 306.083
> > -23.14
> > 40000 1200 454.358 463.215 454.035 457.203 373.345 366.772 357.682
> 365.933
> > -19.96
> > 40000 1300 509.474 501.059 515.199 508.577 390.705 391.746 377.911
> 386.787
> > -23.95
> > 40000 1400 546.603 544.56 554.895 548.686 445.245 445.604 451.151 447.333
> > -18.47
> > 40000 1500 619.412 616.466 610.991 615.623 514.933 496.33 522.703 511.322
> > -16.94
> > 40000 1600 Job failed. Job failed.
> >
> > 2014-09-23 13:16 GMT+09:00 Victor Lee <vv...@gmail.com>:
> >>
> >> Hello all.
> >>
> >> I conducted a perfromance test about latest version 0.7.0 of Apache Hama
> >> (applied Kryo)
> >> (Kryo is a fast and efficient object graph serialization framework for
> >> Java)
> >>
> >> The results of the test, the performance has been greatly
> >> improved(20~25%).
> >> Kryo's serialization API is speed, efficiency and an easy,
> >> so useful for hama's network communication.
> >>
> >> It seems a very good improvement!. : )
> >>
> >> I will attach the test results below.
> >>
> >> Thanks.
> >>
> >>
> >>
> >> ( supported by Byungseok Min. Thanks ! )
> >>
> >>
> >
>
>
>
> --
> Best Regards, Edward J. Yoon
> CEO at DataSayer Co., Ltd.
>
Re: PageRank Benchmark Test (0.6.4 vs 0.7.0-SNAPSHOT)
Posted by "Edward J. Yoon" <ed...@apache.org>.
Yes, I saw it. I personally want to use the OffHeapVerticesInfo
implementation as a default, and clean up all implementations which
doesn't support dynamic graph operations.
On Thu, Sep 25, 2014 at 4:00 PM, Tommaso Teofili
<to...@gmail.com> wrote:
> 2014-09-24 11:16 GMT+02:00 Edward J. Yoon <ed...@apache.org>:
>
>> Thanks for your benchmarks. BTW, I guess I need to explain more about this.
>>
>> As I described HAMA-919, we have to manage the outgoing messages per
>> vertex (like giraph), and won't use sorted queue for graph job
>> anymore. This is the way to improve the performance and memory
>> efficiency. With this, I think VerticesInfo also can be changed. So
>> now, it's a meaningless testing. I kindly ask you to benchmark again
>> after HAMA-919 is done.
>>
>> However, one thing is clear that Kryo serializer is really useful. :-)
>>
>
> by the way, the OffHeapVerticesInfo (based on Apache DirectMemory) uses
> Kryo for serialization :-)
>
> Tommaso
>
>
>>
>> On Tue, Sep 23, 2014 at 2:32 PM, Victor Lee <vv...@gmail.com> wrote:
>> > Hello, All ~
>> >
>> > I didn't send attach file. so I send mail again.
>> >
>> > thanks.
>> >
>> >
>> >
>> >
>> > [ Apache Hama Benchmark Test ]
>> > . PageRank Benchmark Test 0.6.4 vs 0.7.0-SNAPSHOT (2014/09/20)
>> > . Environments
>> > - 12 core (2.66Ghz 6core * 2)
>> > - 32G Memory (4G * 8)
>> > - 2 Node Hama Cluster
>> > - 6 tasks (opt : -Xmx1024m, -Xms2048)
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> > . Memory : -Xmx1024m
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> > vertices Edges
>> > per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
>> > rate(%)
>> > 1th 2nd 3rd average 1th 2nd 3rd Average
>> > (seconds)
>> > 40000 500 178.822 178.688 183.85 180.453 147.711 142.232 132.755 140.899
>> > -21.92
>> > 40000 600 231.884 231.964 226.112 229.987 177.764 174.733 171.714 174.737
>> > -24.02
>> > 40000 700 277.061 276.845 273.928 275.945 219.796 222.792 216.817 219.802
>> > -20.35
>> > 40000 800 Job failed. Job failed.
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> > . Memory : -Xmx2048m
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> >
>> > vertices Edges
>> > per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
>> > rate(%)
>> > 1th 2nd 3rd average 1th 2nd 3rd Average
>> > (seconds)
>> > 40000 500 168.741 171.706 168.701 169.716 132.737 130.850 126.670 130.086
>> > -23.35
>> > 40000 600 207.749 208.045 195.854 203.883 159.744 153.757 156.681 156.727
>> > -23.13
>> > 40000 700 240.848 240.837 232.015 237.900 181.217 186.806 187.221 185.081
>> > -22.20
>> > 40000 800 280.729 274.51 280.754 278.664 210.918 209.112 217.013 212.348
>> > -23.80
>> > 40000 900 320.428 320.683 314.3 318.470 235.184 242.585 238.244 238.671
>> > -25.06
>> > 40000 1000 355.425 361.342 358.36 358.376 276.705 279.91 281.294 279.303
>> > -22.06
>> > 40000 1100 394.008 401.419 399.351 398.259 307.003 304.691 306.556
>> 306.083
>> > -23.14
>> > 40000 1200 454.358 463.215 454.035 457.203 373.345 366.772 357.682
>> 365.933
>> > -19.96
>> > 40000 1300 509.474 501.059 515.199 508.577 390.705 391.746 377.911
>> 386.787
>> > -23.95
>> > 40000 1400 546.603 544.56 554.895 548.686 445.245 445.604 451.151 447.333
>> > -18.47
>> > 40000 1500 619.412 616.466 610.991 615.623 514.933 496.33 522.703 511.322
>> > -16.94
>> > 40000 1600 Job failed. Job failed.
>> >
>> > 2014-09-23 13:16 GMT+09:00 Victor Lee <vv...@gmail.com>:
>> >>
>> >> Hello all.
>> >>
>> >> I conducted a perfromance test about latest version 0.7.0 of Apache Hama
>> >> (applied Kryo)
>> >> (Kryo is a fast and efficient object graph serialization framework for
>> >> Java)
>> >>
>> >> The results of the test, the performance has been greatly
>> >> improved(20~25%).
>> >> Kryo's serialization API is speed, efficiency and an easy,
>> >> so useful for hama's network communication.
>> >>
>> >> It seems a very good improvement!. : )
>> >>
>> >> I will attach the test results below.
>> >>
>> >> Thanks.
>> >>
>> >>
>> >>
>> >> ( supported by Byungseok Min. Thanks ! )
>> >>
>> >>
>> >
>>
>>
>>
>> --
>> Best Regards, Edward J. Yoon
>> CEO at DataSayer Co., Ltd.
>>
--
Best Regards, Edward J. Yoon
CEO at DataSayer Co., Ltd.
Re: PageRank Benchmark Test (0.6.4 vs 0.7.0-SNAPSHOT)
Posted by Tommaso Teofili <to...@gmail.com>.
2014-09-24 11:16 GMT+02:00 Edward J. Yoon <ed...@apache.org>:
> Thanks for your benchmarks. BTW, I guess I need to explain more about this.
>
> As I described HAMA-919, we have to manage the outgoing messages per
> vertex (like giraph), and won't use sorted queue for graph job
> anymore. This is the way to improve the performance and memory
> efficiency. With this, I think VerticesInfo also can be changed. So
> now, it's a meaningless testing. I kindly ask you to benchmark again
> after HAMA-919 is done.
>
> However, one thing is clear that Kryo serializer is really useful. :-)
>
by the way, the OffHeapVerticesInfo (based on Apache DirectMemory) uses
Kryo for serialization :-)
Tommaso
>
> On Tue, Sep 23, 2014 at 2:32 PM, Victor Lee <vv...@gmail.com> wrote:
> > Hello, All ~
> >
> > I didn't send attach file. so I send mail again.
> >
> > thanks.
> >
> >
> >
> >
> > [ Apache Hama Benchmark Test ]
> > . PageRank Benchmark Test 0.6.4 vs 0.7.0-SNAPSHOT (2014/09/20)
> > . Environments
> > - 12 core (2.66Ghz 6core * 2)
> > - 32G Memory (4G * 8)
> > - 2 Node Hama Cluster
> > - 6 tasks (opt : -Xmx1024m, -Xms2048)
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > . Memory : -Xmx1024m
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > vertices Edges
> > per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
> > rate(%)
> > 1th 2nd 3rd average 1th 2nd 3rd Average
> > (seconds)
> > 40000 500 178.822 178.688 183.85 180.453 147.711 142.232 132.755 140.899
> > -21.92
> > 40000 600 231.884 231.964 226.112 229.987 177.764 174.733 171.714 174.737
> > -24.02
> > 40000 700 277.061 276.845 273.928 275.945 219.796 222.792 216.817 219.802
> > -20.35
> > 40000 800 Job failed. Job failed.
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > . Memory : -Xmx2048m
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > vertices Edges
> > per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
> > rate(%)
> > 1th 2nd 3rd average 1th 2nd 3rd Average
> > (seconds)
> > 40000 500 168.741 171.706 168.701 169.716 132.737 130.850 126.670 130.086
> > -23.35
> > 40000 600 207.749 208.045 195.854 203.883 159.744 153.757 156.681 156.727
> > -23.13
> > 40000 700 240.848 240.837 232.015 237.900 181.217 186.806 187.221 185.081
> > -22.20
> > 40000 800 280.729 274.51 280.754 278.664 210.918 209.112 217.013 212.348
> > -23.80
> > 40000 900 320.428 320.683 314.3 318.470 235.184 242.585 238.244 238.671
> > -25.06
> > 40000 1000 355.425 361.342 358.36 358.376 276.705 279.91 281.294 279.303
> > -22.06
> > 40000 1100 394.008 401.419 399.351 398.259 307.003 304.691 306.556
> 306.083
> > -23.14
> > 40000 1200 454.358 463.215 454.035 457.203 373.345 366.772 357.682
> 365.933
> > -19.96
> > 40000 1300 509.474 501.059 515.199 508.577 390.705 391.746 377.911
> 386.787
> > -23.95
> > 40000 1400 546.603 544.56 554.895 548.686 445.245 445.604 451.151 447.333
> > -18.47
> > 40000 1500 619.412 616.466 610.991 615.623 514.933 496.33 522.703 511.322
> > -16.94
> > 40000 1600 Job failed. Job failed.
> >
> > 2014-09-23 13:16 GMT+09:00 Victor Lee <vv...@gmail.com>:
> >>
> >> Hello all.
> >>
> >> I conducted a perfromance test about latest version 0.7.0 of Apache Hama
> >> (applied Kryo)
> >> (Kryo is a fast and efficient object graph serialization framework for
> >> Java)
> >>
> >> The results of the test, the performance has been greatly
> >> improved(20~25%).
> >> Kryo's serialization API is speed, efficiency and an easy,
> >> so useful for hama's network communication.
> >>
> >> It seems a very good improvement!. : )
> >>
> >> I will attach the test results below.
> >>
> >> Thanks.
> >>
> >>
> >>
> >> ( supported by Byungseok Min. Thanks ! )
> >>
> >>
> >
>
>
>
> --
> Best Regards, Edward J. Yoon
> CEO at DataSayer Co., Ltd.
>
Re: PageRank Benchmark Test (0.6.4 vs 0.7.0-SNAPSHOT)
Posted by "Edward J. Yoon" <ed...@apache.org>.
Thanks for your benchmarks. BTW, I guess I need to explain more about this.
As I described HAMA-919, we have to manage the outgoing messages per
vertex (like giraph), and won't use sorted queue for graph job
anymore. This is the way to improve the performance and memory
efficiency. With this, I think VerticesInfo also can be changed. So
now, it's a meaningless testing. I kindly ask you to benchmark again
after HAMA-919 is done.
However, one thing is clear that Kryo serializer is really useful. :-)
On Tue, Sep 23, 2014 at 2:32 PM, Victor Lee <vv...@gmail.com> wrote:
> Hello, All ~
>
> I didn't send attach file. so I send mail again.
>
> thanks.
>
>
>
>
> [ Apache Hama Benchmark Test ]
> . PageRank Benchmark Test 0.6.4 vs 0.7.0-SNAPSHOT (2014/09/20)
> . Environments
> - 12 core (2.66Ghz 6core * 2)
> - 32G Memory (4G * 8)
> - 2 Node Hama Cluster
> - 6 tasks (opt : -Xmx1024m, -Xms2048)
>
>
>
>
>
>
>
>
>
>
>
> . Memory : -Xmx1024m
>
>
>
>
>
>
>
>
>
>
>
> vertices Edges
> per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
> rate(%)
> 1th 2nd 3rd average 1th 2nd 3rd Average
> (seconds)
> 40000 500 178.822 178.688 183.85 180.453 147.711 142.232 132.755 140.899
> -21.92
> 40000 600 231.884 231.964 226.112 229.987 177.764 174.733 171.714 174.737
> -24.02
> 40000 700 277.061 276.845 273.928 275.945 219.796 222.792 216.817 219.802
> -20.35
> 40000 800 Job failed. Job failed.
>
>
>
>
>
>
>
>
>
>
>
> . Memory : -Xmx2048m
>
>
>
>
>
>
>
>
>
>
>
> vertices Edges
> per vertex 0.6.4 0.7.0-SNAPSHOT(2014/09/20) Improvement
> rate(%)
> 1th 2nd 3rd average 1th 2nd 3rd Average
> (seconds)
> 40000 500 168.741 171.706 168.701 169.716 132.737 130.850 126.670 130.086
> -23.35
> 40000 600 207.749 208.045 195.854 203.883 159.744 153.757 156.681 156.727
> -23.13
> 40000 700 240.848 240.837 232.015 237.900 181.217 186.806 187.221 185.081
> -22.20
> 40000 800 280.729 274.51 280.754 278.664 210.918 209.112 217.013 212.348
> -23.80
> 40000 900 320.428 320.683 314.3 318.470 235.184 242.585 238.244 238.671
> -25.06
> 40000 1000 355.425 361.342 358.36 358.376 276.705 279.91 281.294 279.303
> -22.06
> 40000 1100 394.008 401.419 399.351 398.259 307.003 304.691 306.556 306.083
> -23.14
> 40000 1200 454.358 463.215 454.035 457.203 373.345 366.772 357.682 365.933
> -19.96
> 40000 1300 509.474 501.059 515.199 508.577 390.705 391.746 377.911 386.787
> -23.95
> 40000 1400 546.603 544.56 554.895 548.686 445.245 445.604 451.151 447.333
> -18.47
> 40000 1500 619.412 616.466 610.991 615.623 514.933 496.33 522.703 511.322
> -16.94
> 40000 1600 Job failed. Job failed.
>
> 2014-09-23 13:16 GMT+09:00 Victor Lee <vv...@gmail.com>:
>>
>> Hello all.
>>
>> I conducted a perfromance test about latest version 0.7.0 of Apache Hama
>> (applied Kryo)
>> (Kryo is a fast and efficient object graph serialization framework for
>> Java)
>>
>> The results of the test, the performance has been greatly
>> improved(20~25%).
>> Kryo's serialization API is speed, efficiency and an easy,
>> so useful for hama's network communication.
>>
>> It seems a very good improvement!. : )
>>
>> I will attach the test results below.
>>
>> Thanks.
>>
>>
>>
>> ( supported by Byungseok Min. Thanks ! )
>>
>>
>
--
Best Regards, Edward J. Yoon
CEO at DataSayer Co., Ltd.
Re: PageRank Benchmark Test (0.6.4 vs 0.7.0-SNAPSHOT)
Posted by Victor Lee <vv...@gmail.com>.
Hello, All ~
I didn't send attach file. so I send mail again.
thanks.
[ Apache Hama Benchmark Test ]
. PageRank Benchmark Test 0.6.4 vs 0.7.0-SNAPSHOT (2014/09/20)
. Environments
- 12 core (2.66Ghz 6core * 2)
- 32G Memory (4G * 8)
- 2 Node Hama Cluster
- 6 tasks (opt : -Xmx1024m, -Xms2048)
*. Memory : -Xmx1024m*
*vertices*
*Edges per vertex* *0.6.4* *0.7.0-SNAPSHOT(2014/09/20)*
*Improvement rate(%)* *1th* *2nd* *3rd* *average* *1th* *2nd* *3rd*
*Average(seconds)* 40000 500 178.822 178.688 183.85 *180.453* 147.711
142.232 132.755 *140.899* *-21.92* 40000 600 231.884 231.964 226.112
*229.987* 177.764 174.733 171.714 *174.737* *-24.02* 40000 700 277.061
276.845 273.928 *275.945* 219.796 222.792 216.817 *219.802* *-20.35* 40000
800 Job failed. Job failed.
*. Memory : -Xmx2048m*
*vertices*
*Edges per vertex* *0.6.4 * *0.7.0-SNAPSHOT(2014/09/20)*
*Improvement rate(%)* *1th* *2nd* *3rd* *average* *1th* *2nd* *3rd*
*Average(seconds)* 40000 500 168.741 171.706 168.701 *169.716* 132.737
130.850 126.670 *130.086* *-23.35* 40000 600 207.749 208.045 195.854
*203.883* 159.744 153.757 156.681 *156.727* *-23.13* 40000 700 240.848
240.837 232.015 *237.900* 181.217 186.806 187.221 *185.081* *-22.20* 40000
800 280.729 274.51 280.754 *278.664* 210.918 209.112 217.013 *212.348*
*-23.80* 40000 900 320.428 320.683 314.3 *318.470* 235.184 242.585 238.244
*238.671* *-25.06* 40000 1000 355.425 361.342 358.36 *358.376* 276.705
279.91 281.294 *279.303* *-22.06* 40000 1100 394.008 401.419 399.351
*398.259* 307.003 304.691 306.556 *306.083* *-23.14* 40000 1200 454.358
463.215 454.035 *457.203* 373.345 366.772 357.682 *365.933* *-19.96* 40000
1300 509.474 501.059 515.199 *508.577* 390.705 391.746 377.911 *386.787*
*-23.95* 40000 1400 546.603 544.56 554.895 *548.686* 445.245 445.604
451.151 *447.333* *-18.47* 40000 1500 619.412 616.466 610.991 *615.623*
514.933 496.33 522.703 *511.322* *-16.94* 40000 1600 Job failed. Job
failed.
2014-09-23 13:16 GMT+09:00 Victor Lee <vv...@gmail.com>:
> Hello all.
>
> I conducted a perfromance test about latest version 0.7.0 of Apache Hama
> (applied Kryo)
> (Kryo is a fast and efficient object graph serialization framework for
> Java)
>
> The results of the test, the performance has been greatly improved(20~25%).
> Kryo's serialization API is speed, efficiency and an easy,
> so useful for hama's network communication.
>
> It seems a very good improvement!. : )
>
> I will attach the test results below.
>
> Thanks.
>
> [image: 埋め込み画像 1]
>
> ( supported by Byungseok Min. Thanks ! )
>
>
>