Performance Analysis and Optimization of the Hurricane File System ...

CHAPTER 7. MACROBENCHMARK 74
avg # cycles per processor
6e+10
4e+10
2e+10
0e+00
2 4 6 8 10 12
# processors
Unoptimized
Optimized
Optimized 2
Figure 7.4: Web server – further optimized.
avg # cycles per processor
6e+10
4e+10
2e+10
0e+00
2 4 6 8 10 12
# processors
Unoptimized
Optimized
Optimized 2 smaller
Optimized 2
Figure 7.6: Web server – further optimized,
smaller.
avg # cycles per processor
2e+09
1.5e+9
1e+09
5e+08
0e+00
2 4 6 8 10 12
# processors
Optimized 2
Figure 7.5: Web server – further optimized, magnified.
avg # cycles per processor
5e+09
4e+09
3e+09
2e+09
1e+09
0e+00
2 4 6 8 10 12
# processors
Optimized 2 smaller
Optimized 2
Figure 7.7: Web server – further optimized,
smaller, magnified.
threads are accessing the same file, hash collisions may still not occur if they access sufficiently different
locations of the file, since access to the group of 16 hash lists is roughly based on file block number.
The results from these additional optimizations are shown in Figure 7.4 with the curve labeled “Op-
timized 2”. Scalability was greatly improved, however, ideal scalability was not achieved. The resulting
curve, magnified in Figure 7.5, indicates a minor scalability degradation. Average thread execution time
on 12 processors was 67% greater than on a uniprocessor. To briefly examine sensitivity to the number of
hash lists in a grouping, the ORS cache hash list group size was reduced from 64 to 16, and the block cache
hash list group size was reduced from 256 to 64. The results are shown Figure 7.6 and Figure 7.7 6 with
the curve labeled “Optimized 2 smaller”. As expected, the curve was shifted higher, but minor scalability
problems unexpectedly reappeared since the slope of the curve has increased slightly. The modified hash
function guaranteed exclusive hash list access under our particular workload so the scalability problems were
not caused by hash list sharing. We did not investigate these scalability problems further since the smaller
6 64 MB of RAM was used on the dual processor configuration instead of 128 MB due to OS stability problems.