■ THIS WEEK (9/8~9/14)
□ CS 710 project
I compared the cpi of IDEAL, MESH, PT_TO_PT.
Where PT_TO_TO is the topology that garnet offers defaultly (it is almost ideal).
MESH is the 4x4 mesh topology,
IDEAL uses the same link latency and config with MESH but connected in a p2p manner.
For 3 application: barnes, cont-ocean, noncont-ocean,
The cpi result is PT_TO_PT < IDEAL < MESH
But the gap between them is not so big.
[mesh_vs_ideal.xlsx]
TODO: i think that the network topology does not seriously affect to the performance in splash2 applications
We might need to test specjbb or other and compare the results.
□ Booksim
I plotted 3-D graph of accepted packets for each nodes, but couldn’t find the specific pattern.
[booksim_hotspot.xlsx]
TODO: I want to fix the trafficmanager code to inject the same amount of packets for voq and no-voq.
and then, compare the results.
□ Phoenix
I changed the wordcount_map function to count all words.
(before, it only counted the words which consist of a-z)
The speedup pattern of map phase is not changed,
But the reduce time hugely increases.
From the result, the reduce processing time depends on the amount of words.
Actually,
For the mapreduce version of wordcount, map threads store all words to the intermediate queue,
after that, reduce threads read words from the intermediate queue.
For the pthreads version, there is no intermediate queue between map and reduce.
So, every word passes map and reduce at a time.
[test_on_gems.xlsx]
TODO: I think I need to check the size of intermediate queue,
and check the latency of store, read words.
