下一个运行阶段是 S5GC1

Youth

http://einstein.phys.uwm.edu/forum_thread.php?id=7811

Posted 3 May 2010 22:20:27 UTC - in response to Message 103851.

With S5GCE over 90% done, I was wondering what would be next. Is there is enough S6 data to start crunching that?

S5GCE的完成度已经90%啦。

Thanks for reminding me of an update. There is quite some work in progress and discussion going on right now as we prepare the next run called "S5GC1".

下一运行阶段是S5GC1，去掉了GCE里面的E，也就是Engineering。

We'll use the same data that we did for S5GCE - the best 25h stretches from both years of S5. We'll have applications that run about 3x as fast as the current ones, there's a bit room left for further speed improvements. We'll use the same FStat code that we've been using more or less unchanged since S5R1; the new "resampling FStat" code isn't ready to be used yet. S5GC1 will be designed for a total runtime of 6 months.

使用的还是GCE的数据，也就是S5中得到的最好的25小时的数据。新的计算程序将比现在的程序快2倍，而且还有进一步优化的空间。FStat代码在S5R1之后没啥大变化，而新的resampling FStat暂时还用不上。GC1的运行时间预计在6个月。

We're currently discussing measurements to not stress our download servers (and the client wires) by the application speedup. One thing that would be astrophysical interesting and easy to implement is to increase the spindown range we search for in every workunit by about the factor of the application speedup. This would mean that although a GC1 task does three times the computation of a GCE one, its runtime and therefore the total computing time per download volume ratio will stay about the same. We're also discussing then to cut the resulting workunits in halve, they are a bit on the longer end now.

我们正在考虑如果避免增加下载服务器的负荷（考虑到新计算程序的计算速度提升），最有可能的估计还是增加单个计算任务的计算量。

BM

liqing

好啊。我的主力计算项目，觉得这个有意义罢了。

Youth

Posted 4 May 2010 7:18:21 UTC - in response to Message 103857.

Are there plans to deploy the new FStat code midway through S5GC1, or will it result in differences large enough to require waiting until the next run?

There are no solid plans to do so. We may test applications with that code during S5GC1, but to unveil the codes full speedup potential one needs to set up the workunits differently and possibly generate new data files, too, which would mean a new run. However if things go surprisingly well we may start the next run well before S5GC1 ends.

BM

刚看了下之前GCE的新闻，有点搞不明白了，按理GCE就是在测试新的FStat代码，而如果GC1还在使用老的FStat代码，是不是还在继续做测试？

cnchina

早上打开boinc，看到收到这一批包了,还没开始算
计算程序是Global Correlations S5 Search #1 3.02
原先的是Global Correlations S5 Engineering 3.04 (S5GCESSSE2)

Youth

嗯，今天开始发放GC1。

Youth

我机器上也有一个，现在内存占用有200兆咯~~

cnchina

回复 6# Youth


    S5GCE不就已经180兆了吗?

前朝遗少

11504 root      39  19  100 189:02.19  2.4  196m 192m 1148 R einstein_S5GCE_

Youth

回7楼：嗯，只是增加了一点点：）

回8楼：好像win版的gce内存占用稍小一些？

cnchina

跑完一个包，发觉时间比S5GCE长了25%，分数却没有加。典型的加料不加价呀……

socratism

感觉是不是原始数据或者后续数据的整理速度跟不上分布式计算的速度了？

Youth

还好，至少还有东西可以算，没东西算的时候才是跟不上了：）

liqing

现在我收到的是 3.02和3.08的包。

Youth

3.08的是ABP子项目

liqing

3.08的是ABP子项目
Youth 发表于 2010-5-28 13:40

ABP项目是啥东西？