SRPT Scheduling in Web Servers

Amod Jaltade and Aditya Jaltade 
18-845, Internet Services, Spring 2014 

Today’s web-servers are very busy and handle a very large number of
requests. Consequently, their performance under non-trivial load
conditions is very important. One of the metrics that can be used to
measure server performance is “response-time." The response time of a
request depends on a lot of factors, such as server load and
scheduling policies.  One way to decrease the response times is to let
the webserver schedule these requests intelligently.

The SYNC project at Carnegie Mellon studied the effect of scheduling
requests inside a web-server. Their research claims that scheduling
requests for static content using the
Shortest-Remaining-Processing-Time first algorithm reduces the mean
response time by a factor 2 to 5 under loads greater than 50%, without
any significant effect on larger requests.  

Our study is aimed at verifying these claims. We have implemented the
SRPT scheduling mechanism for the Linux kernel. Our implementation
consists of a kernel module and an userspace library, and requires
minimal changes to the web-server’s code. When enabled, our mechanism
ensures that processing of packets for requests with fewer
remainingbytes (to be sent) is prioritized over processing for the
ones with more remaining-bytes. Our experiments with the tiny
web-server indicate that mean response times reduce by a factor of 2
to 3 when the load is greater than 50%, for workloads that follow a
heavy-tailed distribution.