Weekly report (9/28)

Hi. 

First of all, I apologize for late report every week :'( 

I will be more punctual next time. 

@ mapreduce ppt 

The PPT is almost done except for the summary part. 

I thought about what to write, but I think I need to discuss about the new results. 

I spent time to explain the mapreduce concept first, 

And compared the result in the paper with my test result. (ver1.0 and ver2.0 on sunT2 machine) 

Then, I added the dynamic load balancing and execution time breakdown. 

And I divided the test configurations as 4 part, which are input size, thread distribution, cache size, and page size. 

Actually, I tested page size this week, and found that: 

When I changed the page size to 64K (default is 8K), there was no degradation for wordcount. 

So I am testing if I change the page size, 

whether the degradation will disappear on simics result. (where there was degradation after 16 nodes) 

@ Booksim 

I printed the buffer size for each cycle, and found that the buffers of voq were saturated much faster than no-voq. 

I thought that it was because the total amount of packets of voq is more than no-voq. 

So I printed the latency of each result only whey the cycle=5000, 2000. 

then, the voq finally printed out better result than no-voq. 

(and I checked that voq is better than no-voq for the same cycles. 

But final # cycle of voq is twice more than no-voq because of the # of packets.) 

I will test with varying buf size and compare the results. 

Thanks. 

minjeong