Comparison with Other Caching Policies

Next: Empirical Sensitivity Analysis of Up: Simulation Results Previous: Simulation Results

Comparison with Other Caching Policies

The simulation results on the W3C Web server are illustrated in Figures 1, 2. Note that the curves correspond to the miss rates (and not on hit rates) of a day. The variation in time of the efficiency of the caching (removal) policies will be discussed later.

**Figure 1:** *W3C server - request miss rate*
$\begin{figure} \centerline{ \psfig {figure=W3C_miss.ps,width=10cm,angle=-90} }\end{figure}$

**Figure 2:** *W3C server - byte miss rate*
$\begin{figure} \centerline{ \psfig {figure=W3C_bytemiss.ps,width=10cm,angle=-90} }\end{figure}$

As already observed in [8] and mentioned previously, very high hit rates (both for requests and bytes) are obtained with a small cache size. With only 256KB, the request hit rate ranges from 55% to 75%, and the byte hit rate is around 40% (except for SIZE, which is only 15%). With a cache of 8MB, the request hit rate is more than 90% with all the policies, and the byte hit rate around 85%, except for the SIZE policy (60%).

It is clear from the comparison, that static caching outperforms all of the others in terms of byte hit rate for any trace and any cache size. It is also better than all of the removal policies (except SIZE) in terms of reducing the request hit rate.

The simulation results confirm our remarks about the behavior of the SIZE policy. It is good for the request hit rate (for cache size > 8MB), but it performs poorly for the byte hit rate (60%, whereas the others have byte hit rate greater than 80%).

Among all the policies other than SIZE, LRU is the worst policy in all cases. LFU is quite similar to static, but slightly less efficient. For example, a 4MB cache with a static policy will read around 100MB/day less from disk than a cache with LFU, and almost 250MB/day less than LRU.

The simulation results on the INRIA Web server are illustrated in Figures 3, 4. Note that, again, the curves correspond to the miss rates instead of hit rates.

**Figure 3:** *INRIA server - request miss rate*
$\begin{figure} \centerline{ \psfig {figure=INRIA_miss.ps,width=10cm,angle=-90} }\end{figure}$

**Figure 4:** *INRIA server - byte miss rate*
$\begin{figure} \centerline{ \psfig {figure=INRIA_bytemiss.ps,width=10cm,angle=-90} }\end{figure}$

One can see that observations made previously about the efficiency of the removal policies on W3C server still holds for the INRIA Web server. However, the miss rates are now higher (for the INRIA server) than in the previous case (the W3C server). With a cache of 8MB, the byte hit rate is around 50%, and the request hit rate around 70%. This phenomenon can be explained by the fact that the most required documents are more concentrated in the W3C Web server than in the INRIA Web server.

Next: Empirical Sensitivity Analysis of Up: Simulation Results Previous: Simulation Results

Nicolas Niclausse
8/25/1997