Check out the new USENIX Web site.

next up previous
Next: Sensitivity Analysis Up: Performance Evaluation Previous: Performance Measurements: Single Applications


Performance Measurements: Multiple Applications


In real systems, multiple applications execute concurrently competing for limited buffer space. To test the DEAR scheme in such an environment, we ran several combinations of two or more of the applications with a buffer cache of 6MB and measured the total number of disk I/Os and the overall response time for both the DEAR scheme and the LRU scheme in FreeBSD. Again, we set the length of the detection period to 500 and the number of sublists in the ordered lists to 5.

 

\psfig{figure=fig/m-io.eps,width=8.35cm} \psfig{figure=fig/m-time.eps,width=8.35cm}

(a) Number of Disk I/Os                              (b) Overall response time

Figure 9: Multiple application performance.

The results in Figure 9 show that the number of disk I/Os is reduced by up to 20% (for the cscope+sort+link case) with an average of 12% and the overall response time by up to 18% (for the glimpse+link case) with an average of 8%.

In the multiple application case, there are two possible benefits from using the proposed DEAR scheme. The first is from applying different replacement policies to different applications based on their detected reference patterns. The second is from giving preference to blocks that belong to an application with the sequential reference pattern when a replacement is needed. To quantify these benefits, we performed an experiment where even the LRU replacement policy gives preference to blocks belonging to an application with the sequential reference pattern, which we call the LRU-SEQ replacement policy.

 

Table 2: Performance comparison between the LRU-SEQ and the DEAR schemes.
Scheme Response Time (seconds)
  cs+sort gli+link cs+wc gli+wc
LRU 70.96 89.87 81.27 89.97
LRU-SEQ 70.72 87.24 71.75 86.88
DEAR 66.61 74.29 62.88 82.36


Table 2 shows the results of the LRU-SEQ scheme for the 6MB buffer cache size. In the case of cscope+sort and glimpse+link, there is little difference between the LRU and the LRU-SEQ schemes, since the reference pattern of the four component applications is not sequential in the steady state. Wc is a utility that displays the numbers of lines, words, and characters in a file. Its steady state reference pattern is sequential. Replacing sort and link with wc, produces a significant difference in the response time between the LRU and the LRU-SEQ schemes. This results from the LRU-SEQ scheme allocating more buffer space to cscope (or glimpse) by replacing blocks of the wc application earlier than the usual LRU order. Still, there is a substantial difference in the response time between the LRU-SEQ scheme and the DEAR scheme indicating that the benefit from applying different replacement polices tailored for different applications is significant.


next up previous
Next: Sensitivity Analysis Up: Performance Evaluation Previous: Performance Measurements: Single Applications

Jongmoo Choi
1999-04-22