6. Related Work

 

Internet caching has attracted a substantial amount of work. Web proxy caches, e.g. the original CERN web cache [14], are client-oriented, while hierarchical internet caches like Harvest/Squid [3] aim at reducing both backbone traffic and end-to-end network latency. As Kroeger et al. [12] and Duska et al. [7] report, Web hit-rates on a wide-area network are only about 40% to 50%; latency can be reduced by 30% to 60%.

We use caching for a completely different purpose. Instead of reducing wide-area traffic and network latency, we aim at improving server latency and server throughput. In a way, that is similar to Iyengar and Challenger [8] who concentrate on how to use caches on a server to improve the generation of dynamic Web pages.

Server operating systems have been discussed recently by Kaashoek et al. [10]. Cheetah's design exploits the underlying characteristics of the Exokernel [11] to construct servers that can access the hardware at low overheads. Cheetah uses kernel extensions to achieve high performance. Our hit-server runs entirely at user level. In some sense, our active objects are similar to ASHes [20] but are geared to multimedia documents, provide an object-oriented structuring technique, run entirely at user level and use supervised IPC.

The facilities we used to securely execute the custom methods of active objects could be used securely support ActiveIP [21]. In this sense, our techniques could be used to build an extremely fast router. One key difference between activeIP and our techniques is that since we use hardware based protection and fast authorized IPC, our code does not need to be interpreted to be supervised but regular binaries can be used instead.

ADC [5] and Fbufs[6] are techniques used to improve the performance of network protocols and drivers for high speed networks and focus on reducing the number of data copies. We use similar techniques but go beyond them as we concentrate more on server throughput and network scheduling, rather than point-to-point network protocol throughput.