In our second set of simulations, we pick the algorithm configuration (1, 1, 0.01) and study how its performance is affected by smaller history structures. The size of the history structure is constrained by specifying a maximum number of nodes implementing an LRU replacement. This size constraint is in addition to the normal aging. The (1, 1, 0.01) configuration is chosen because it only needs two-level trees and yet yields almost the highest latency reduction. Table 1 shows the results on request savings and latency reduction. The 32MB history size is sufficient that nodes are never prematurely removed.
History Size | Request Savings | Latency Reduction |
32MB | 15% | 9% |
4MB | 12% | 7% |
1MB | 9% | 5% |
The results show that when the history size is reduced, the latency reductions decreases, as the algorithm no longer has access to some patterns observed in the past. However, the degradation is gradual and as more memory can be devoted to the history structure, the performance improves.