Evaluating the impact of the API is heavily dependent on the quality of the underlying platform. A slow proxy will mask the overhead of the API, while a fast one will more easily expose the additional latency resulting from the API. The iMimic DataReactor Proxy Cache is a commercial high-performance proxy cache. It supports the standard caching modes (forward, reverse, transparent) and is portable, with versions for the x86, Alpha, and Sparc architectures and the FreeBSD, Linux, and Solaris operating systems. It has performed well in the vendor-neutral Proxy Cache-Offs, setting records in performance, latency, and price/performance [15,16,17].
We test forward-proxy (client-side) cache performance using the Web Polygraph benchmark program running a modified version of the Polymix-3 workload [16]. We use this test and workload because it has the highest number of independently-measured entries of any web proxy benchmark, and it heavily stresses proxy server performance. For the sake of time, we shorten our performance tests to use a 2 hour load plateau instead of four hours, and fill the disks only once before all tests rather than before each test. These changes shorten the load phase of the Polygraph test to roughly 6 hours instead of 10.5, and avoiding a separate fill phase reduces the length of each test by an additional 10-14 hours. The primary performance impact is a 3-4% higher hit ratio than an official test because of a smaller working set and data set. We call this modified test PMix-3.
Polygraph stresses various aspects of proxy performance, particularly in network and disk-related areas. It uses per-connection and per-packet delays between the proxy and simulated remote servers to cause cache misses to have a high response time. Likewise, it generates data sets and working sets that far exceed physical memory, causing heavy disk access. Polygraph stresses connection management by scaling the number of simulated clients and servers with the request rate.
The test system runs FreeBSD 4.4 and includes
a 1400 MHz Athlon, 2 GB of memory, a Netgear GA-620
Gigabit Ethernet NIC, and five 36 GByte 10000 RPM SCSI disks.
All tests
use a target request rate of 1450
.
This throughput
compares favorably with other high-end commercial proxy servers and is over a
factor of ten higher than what free software has
demonstrated [16]. At this rate, the proxy is
managing over 16000 simultaneous connections and 3600 client IP addresses.
Given the fixed request rate, this test
demonstrates any latency differences in the various test scenarios.
(Polygraph also shows some run-to-run randomness in the offered
workload, leading to additional minor variations.)
|
