Information Technology Reference
In-Depth Information
Figure 12. Five interleaved sequential streams
readahead window steadily and is not disturbed
by the random reads.
24.95MB/s for legacy readahead and 28.32MB/s
for demand readahead. When there are 2 and 3
interleaved streams, the legacy readahead quickly
slows down to 7.05MB/s and 3.15MB/s, while
the new one is not affected. When there are 10
interleaved streams, the legacy readahead through-
put crawls along at 0.81MB/s, while the demand
readahead still maintains a high throughput of
21.78MB/s, which is 26.9 times the former one.
We also measured performance with disk
readahead enabled. As showed in figure 13(b),
the in-disk readahead have very positive effects
on the performance. However due to limited disk
cache and capability, it can only support a limited
number of streams. As the number of streams
increase, the influence of the in-disk readahead
decreases. When there are 10 streams, the through-
interleaved Sequential Streams
To validate readahead performance on concur-
rent streams, we created 10 sequential streams
S i , i=1,2,…,10, where Si i is a sequence of 4KB
reads that start from byte (i-1)*100MB and stop
at i*100MB. By interleaving the first n streams,
we get C n =interleave(S 1 ,S 2 ,…,S n ), n=1,2,…,10.
Where C 1 =S 1 is a time and space consecutive read
stream, and C n , n=2,3,…,10 is an interleaved ac-
cess pattern that has n sequential streams. Figure
12 shows a segment of C 5 .
Figure 13 plots the I/O throughputs for C 1 -C 10 .
The single stream throughputs are close, with
Figure 13 Comparison of readahead performances on interleaved reads
Search WWH ::




Custom Search