Results

Table 4 shows the hit ratios of the storage buffer cache with the MQ and LRU replacement algorithms. The difference between the implementation and simulation results is less than 10%, which validates our simulation study. The small difference is mainly caused by two reasons. The first is that the timing is different in the real system due to concurrency. The second is the interaction between cache hit ratios and request rates. When the cache hit ratio increases, the average access time decreases. As a result, more I/O requests are forwarded to the storage system.

Table 4: Percentage hit ratios of the storage buffer cache. The Oracle buffer cache (first level buffer cache) size is always 128 Mbytes.

Storage cache size	MQ	LRU
128MB	19.85	8.85
256MB	31.42	17.66
512MB	44.34	31.69

As shown on Table 4, MQ achieves much higher hit ratios than LRU. For a 512 MBytes storage buffer cache, MQ has a 12.65% higher hit ratio than LRU. Since miss penalty dominates the average access time, we use the relative miss ratio difference to estimate the upper bound of MQ's improvement on the end performance. For a 512 MBytes buffer cache, the relative miss ratio difference between MQ and LRU is 18.5% (

). Therefore, the upper bound for end performance improvement with MQ over LRU is 18.5%.

In fact, in order for LRU to achieve the same hit ratio as MQ, its cache size needs to be doubled. The hit ratio of MQ with a 128 MBytes cache is slightly greater than that of LRU with a 256 MBytes cache. The hit ratio of MQ with a 256 MBytes cache is about the same as LRU with a 512 MBytes cache.

**Figure 7:** Normalized TPC-C transaction rate with different storage buffer cache sizes (All numbers are normalized to the transaction rate achieved by LRU with an 129 MBytes storage buffer cache).
$\begin{figure} \centerline {\psfig{figure=issperf.eps,width=3in}}\end{figure}$

Figure 7 shows the end performance of the MQ and LRU algorithms. For all three buffer cache sizes, MQ improves the TPC-C end performance over LRU by 8-11%. Due to certain license problems, we are not allowed to report the absolute performance in terms of transaction rate. Therefore, all performance numbers are normalized to the transaction rate with a 128 MBytes buffer cache using the LRU replacement algorithm. Because of the high hit ratios, the MQ algorithm improves the transaction rates over LRU by 8%, 12%, and 10% for 128 MBytes, 256 MBytes and 512 MBytes cache sizes respectively.

Similar to the cache hit ratio improvement, using the MQ algorithm is equivalent to using LRU with a double sized cache. With a 128 MBytes buffer cache, MQ increases the transaction rate by 8%, which is exactly same as the improvement achieved by LRU with a 256 MBytes buffer cache. MQ with a 256 MBytes cache achieves a similar transaction rate to LRU with a 512 MBytes cache.