Check out the new USENIX Web site. next up previous
Next: File system limitations Up: Storage management for web Previous: Abstract


Caching web proxies are computer systems dedicated to caching and delivering web content. Typically, they exist on a corporate firewall or at the point where an Internet Service Provider (ISP) peers with its network access provider. From a web performance and scalability point of view, these systems have three purposes: improve web client latency, drive down the ISP's network access costs because of reduced bandwidth requirements, and reduce request load on origin servers.

Squid and Apache are two popular web proxies. Both of these systems use the standard file system services provided by the host operating system. On UNIX this is usually UFS, a descendant of the 4.2BSD UNIX Fast File System (FFS) [13]. FFS was designed for workstation workloads and is not optimized for the different workload and requirements of a web proxy. It has been observed that file system latency is a key component in the latency observed by web clients [21].

Some commercial vendors have improved I/O performance by rebuilding the entire system stack: a special operating system with an application-specific file system executing on dedicated hardware (e.g., CacheFlow, Network Appliance). Needless to say, these solutions are expensive. We believe that a lightweight and portable file system can be built that will allow proxies to achieve performance close to that of a specialized system on commodity hardware, within a general-purpose operating system, and with minimal changes to their source code; Gabber and Shriver [5] discuss this view in detail.

We have built a simple, lightweight file system library named Hummingbird that runs on top of a raw disk partition. This system is easily portable--we have run it with minimal changes on FreeBSD, IRIX, Solaris, and Linux. In this paper we describe the design, interface, and implementation of this system along with some experimental results that compare the performance of our system with a UNIX file system. Our results indicate that Hummingbird's throughput is 2.3-4.0 times larger than a simulated version of Squid running UFS mounted asynchronously on FreeBSD, 5.4-9.4 times faster than Squid running UFS mounted synchronously on FreeBSD, 5.6-8.4 times larger than a simulated version of Squid running UFS with soft updates on FreeBSD, and 5.4-13 times larger than XFS and EFS on IRIX (see Section 4). We also performed experiments using the Polygraph environment [18] with an Apache proxy; the mean response time for hits in the proxy is 14 times smaller with Hummingbird than with UFS (see Section 5).

Throughout the rest of this paper, we use the terms proxy or web proxy to mean caching web proxy. Section 2 presents the important characteristics of the proxy workload considered for our file system. It also presents some background on file systems and proxies that is important because it motivates much of our design. Section 3 describes the Hummingbird file system. Our experiments and results are presented in Sections 4 and 5. Section 6 discusses related work in file systems and web proxy caching.

next up previous
Next: File system limitations Up: Storage management for web Previous: Abstract
Liddy Shriver 2001-05-01