minimum superstep time

[Yi Lu]

I use the WCC code in the examples folder and run the connected components algorithm on a road-network. There are 3774768 vertices and 16522438 edges, the max degree is 9. Therefore, during the computation, there are not so many msgs need to transmit across the machines. However, I notice the minimum superstep time is around 3 seconds, even though there are less than 100 msgs need to transmit. Overall, GPS needs 17502.681 seconds to run CC on my graph. I also run the algorithm in graphlab, it only needs 2982.3 to complete the computation. I also notice the minimum superstep time in Giraph is also high, about 2~3 seconds, Giraph finishes the computation in 9518.985s.

I wonder whether there is some constant overhead during each superstep, is this a common problem for JAVA implementation of pregel? 

Thanks a lot. Have a good time in your trip.

[Semih Salihoglu]

Yes there's a fix overhead to each superstep, which is set in the polling time. You can  minimize that by setting it to, say 10 milliseconds by using the "-pt 10" option when you call ./start_gps_nodes.sh.

--
You received this message because you are subscribed to the Google Groups "stanfordgpsusers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to stanfordgpsuse...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Yi Lu]

Thanks for your answer. I find the default value is 1000ms, which is much smaller than 3 seconds. By the way, what is the purpose to set the polling time to a large value (e.g. 1000 ms)?

在 2014年3月21日星期五UTC+8下午1时28分35秒，Semih Salihoglu写道：