lmbench

This is a "new idea,implementation and evaluation" paper. The 
designers of lmbench have used a simple principle: a system whose 
basic bulding blocks do not function well cannot work well overall. 
Application of this principle makes lmbench a useful testing tool
for system developers and a valuable benchmark suite for potential
system users. This is not to say that a suite of microbenchmarks alone
is enough to evaluate the system, but it can be very helpful when
results of macrobenchmarks are untrustworthy or complicated.

lmbench has been carefully designed: the issues of sizing the
benchmark, compile time issues and multiprocessor issues have been
considered. The tests are very easy to interpret: For example the
memory bandwidth test reports results from simple memory reads 
along with performance of bcopy - a routine very commonly used
in the system. The presence of these simpler tests makes it easy 
to predict the expected behavior and detect anomalies.

The goal of lmbench was to target the basic building blocks of the
system. The important question to answer, therefore, is 
whether the lmbench designers have correctly and completely 
identified these basic building blocks. In my opinion, the lmbench
could be improved if it did the following:

1. Many tests measure performance of software and hardware together, 
while it would be nice to see pure software overhead. lmbench results 
for several OS's running on the same hardware could help to isolate
the overhead of the software.

2. Virtual memory is used so much in systems nowadays that it could
be considered as a system's basic building block. It would be interesting
to measure the vm system overhead.

Overall, the paper is very well written, clear and easy to follow and 
provides significant contribution to research community.