Internet information systems have evolved so rapidly that they postponed performance and scalability for the sake of functionality and easy deployment. However, they cannot continue to meet exponentially growing demand without new infrastructure. Towards this end, we designed the Harvest hierarchical object cache.
This paper presents measurements that show that the Harvest cache achieves better than an order of magnitude performance improvement over other proxy caches. It also demonstrates that HTTP is not an inherently slow protocol, but rather that many popular implementations have ignored the sage advice to make the common case fast [14].
Hierarchical caching distributes load away from server hot spots raised by globally popular information objects, reduces access latency, and protects the network from erroneous clients. High performance is particularly important for higher levels in the cache hierarchy, which may experience heavy service request rates.
The Internet's autonomy and scale present difficult challenges to the way we design and build system software. Once software becomes accepted as de facto standards, both its merits and its deficiencies may live forever. For this reason, the real-world complexities of the Internet make one face difficult design decisions. The maze of protocols, independent software implementations, and well-known bugs that comprise the Internet's upper layers, frequently force tradeoffs between design cleanliness and operational transparency. This paper discusses many of the tradeoffs forced upon us.