Check out the new USENIX Web site. next up previous
Next: Results Up: Simulation and Results Previous: Simulation and Results

Simulation Experiments

We have implemented the nine replacement algorithms in our buffer cache simulator. The block size for all simulations is 8 KBytes. With experiements, we found out that using $log(f)$ function as our $QueueNum(f)$ function works very well for all traces. Our experiments also show that eight LRU queues are enough to seperate the warmer blocks from the others. The history buffer $Q_{out}$ size is set to be four times of the number of blocks in the cache. Each entry of the history buffer occupies fewer than 32 bytes so that the memory requirement for the history buffer is quite small, less than 0.5% of the buffer cache size. The lifeTime parameter is adjusted adaptively at running time. The main idea for dynamic lifeTime adjusting is to efficiently collect statistic information on the temporal distance distributions from access history. Due to page limits, we will not discuss it in this paper, but details can be found in  [44]. The history buffer size for 2Q is one half of the number of blocks in the cache as suggested by Johnson and Shasha in [23]. For fairness, we have extended the LFRU, LRU-2, FBR and LFU algorithms to use a history buffer to keep track of $CRF$ values, second-to-last reference time and access frequencies for recently evicted blocks respectively (see section 2), using a history buffer of size equal to that in MQ. We have tuned the FBR and LFRU algorithms with several different parameters as suggested by the authors and report the best performance. The off-line optimal algorithm (OPT) was first proposed by Belady [2,17] and is widely used to derive the lower bound of cache miss ratio for a given reference string. This algorithm replaces the block with the longest future reference distance. Since it relies on the precise knowledge of future references, it cannot be used on-line. ws Belady's OPT algorithm and WORST algorithm [2,17] As with all cache studies, interesting effects can only be observed if the size of the working set exceeds the cache capacity. The three traces provided by other sources (HP Disk Trace, Auspex Server Trace and Web Server Trace) have relatively small working sets. To anticipate the current trends that working set sizes increase with user demands and new technologies, we chose to use smaller buffer cache sizes for these three traces. In most of experiments, we set the second level buffer cache size to be larger than the first level buffer cache size. However, this property does not always hold in real systems, For example, most of storage systems such as the IBM Enterprise Storage Server have less than 1 Giga Bytes of storage cache (second level buffer cache), while the frontier server, database or file servers, typically have more than 2 Gigabytes of buffer cache (first level buffer cache). Because of this reason, we have also explored a few cases where the second level buffer cache is equal to or smaller than the first level buffer cache.

next up previous
Next: Results Up: Simulation and Results Previous: Simulation and Results
Yuanyuan Zhou
2001-04-29