Check out the new USENIX Web site. next up previous
Next: Effect of request rate Up: Scalable kernel performance for Previous: Scalability with respect to

Performance of a live system


Digital Equipment Corporation operates a Web proxy system, in Palo Alto, California, that serves a large fraction of Digital's internal users. During a typical weekday, the system handles as many as 2.6 million HTTP requests, from at least 5570 individual client hosts.

We installed our modified kernel on the proxy server, a 500 MHz AlphaStation 500 system (21164A processor, SPECInt95 = 15.0) with 512 MBytes of RAM. We then ran the system using either the unmodified kernel or our modified kernel, each for an entire calendar day (midnight to midnight, Pacific Time), and collected extensive monitoring information.

During these these tests, the proxy server used version 3.1.2c-OSF of the NetCache software [16] from Network Appliance, Inc. Like Squid, NetCache was based on the Harvest Cache software, although NetCache and Squid have since evolved separately. Because caching tends to reduce the number of simultaneous network connections, during our trials we operated this software with caching disabled. This increases the load on the system, but for various reasons does not significantly increase response time as seen by the users.

Table 3: Statistics for live tests

Table 3 shows some statistics for each of the trials. The ``Max. alloc. fds'' column shows the largest number of file descriptors allocated to a single process at any one point during the trial; the ``Peak req. rate'' column shows the largest number of requests logged during a single second over the course of the day.

Gaurav Banga
Mon Apr 27 13:10:55 CDT 1998