Cache Size

Tables 3 and 4 compare the performance of the DEAR scheme against the LRU scheme for various buffer cache sizes for the single and multiple application cases, respectively. The results from the single application case show that as long as the total number of distinct blocks accessed by an application is greater than the number of blocks in the buffer cache, there is a substantial difference in the response time between the DEAR and the LRU schemes. However, when the number of distinct blocks of an application is smaller than the number of blocks in the buffer cache, all the blocks are cached in the buffer cache and the two schemes show similar performance. The latter behavior is most visible for the link application that has the smallest number of distinct blocks (about 310 blocks). For link, the DEAR and the LRU schemes provide similar response times.

 

Table 3: Single application performance for various buffer cache sizes.

Application	Scheme	Response Time (seconds)
		2MB	4MB	6MB	8MB
cscope	DEAR	16.99	14.90	12.87	11.17
	LRU	19.79	19.79	19.77	19.77
glimpse	DEAR	39.12	35.68	33.73	32.87
	LRU	40.70	39.72	39.55	37.49
sort	DEAR	18.16	15.54	13.60	12.10
	LRU	18.50	17.45	16.59	14.68
link	DEAR	28.19	23.38	23.38	23.38
	LRU	29.65	23.35	23.35	23.35
cpp	DEAR	132.94	94.42	91.61	91.36
	LRU	159.59	97.94	93.39	91.82
gnuplot	DEAR	43.54	42.26	41.39	41.19
	LRU	44.30	44.30	44.30	44.30
postgres1	DEAR	38.37	36.16	34.22	32.17
	LRU	39.72	38.91	38.82	38.76
postgres2	DEAR	74.57	72.51	71.15	68.45
	LRU	82.93	74.93	74.75	73.93

 

Table 4: Multiple application performance for various buffer cache sizes.

Applications	Scheme	Response Time (seconds)
		4MB	6MB	8MB	10MB
cs+sort	DEAR	70.4	66.6	62.9	53.5
	LRU	71.5	70.9	69.9	67.3
gli+link	DEAR	79.1	74.2	71.6	70.1
	LRU	94.5	89.8	79.1	77.9
cpp+ps1	DEAR	222.2	216.7	209.8	202.4
	LRU	236.5	229.9	226.6	226.1
gli+ps2	DEAR	145.6	139.8	132.5	128.3
	LRU	165.5	155.2	146.7	138.8
cs+sort+link	DEAR	116.1	112.7	106.7	101.8
	LRU	121.3	118.0	112.8	105.3
gli+sort+cpp	DEAR	245.9	235.5	215.8	207.2
	LRU	246.3	245.3	225.9	222.9

 

For the multiple application case, the case where the total number of distinct blocks accessed by the component applications is smaller than the number of blocks in the buffer cache does not occur and the DEAR scheme shows consistently better performance than the LRU scheme.