Performance Metrics

  We are primarily interested in the following performance metrics for our prediction algorithms:

• Latency Reduction: the percentage reduction in the user-visible latency from the prepush technique. The latency reduction comes from two sources:
  • latency hidden: the user-visible latency that is either avoided because the requested document is already in the prefetch buffer (completely hidden), or reduced because it is being pre-pushed (partially hidden).
  • contention avoidance: the reduction in user latency that is due to the document being transfered to the user earlier than in the no-prefetching case. When multiple documents are being sent to the user, the transfers share the limited modem bandwidth. If prefetching can make one of the document transfers happen earlier, then the rest of the document transfers all complete quicker and result in reduction in user-visible latency.

• Wasted Bandwidth: the amount of bytes that are pre-pushed from the proxy to the client but are not read by the client. That is, it is the sum of the sizes of files that are pre-pushed from the proxy to the user but are replaced from the user-side prefetch buffer without ever being accessed by the user. In our following figures we show the ratio between bandwidth wasted and the total bytes accessed by the users.
• Request Savings: the number of time that the user requests a document that is already in the prefetch buffer, or is being sent from proxy to the user. We can further characterize the request savings into the following three categories:
  • prefetched: the document is in the prefetch buffer, and and the user is accessing it for the first time since it is in the prefetch buffer.
  • cached: the document is in the prefetch buffer, but this is not the first time that the uses has accesses it since it has been in the prefetch buffer.
  • partially prefetched: the document is still being sent from the proxy to the user.
  The reason for separating the request savings into the three categories is that we would like to separate the caching effect of the prefetch buffer from the prefetching effect. The prefetch buffer also acts as an extended browser cache for the user, and to understand the latency reduction due to the prediction algorithm, it is important to separate the two effects.

Latency reduction is the primary goal of the pre-push scheme. Bandwidth wasted measure the extra bandwidth consumed by the algorithm. Unlike in the wide-area network, wasted bandwidth on the modem line can be tolerated: the modem line would stay idle anyhow. For most users who do not initiate other network transfers (such as ftp) while Web-surfing, the wasted bandwidth has virtually no effect. Finally, we need to investigate request savings to understand the source of latency reduction.