Phoronix benchmarks DF 3.2 vs Ubuntu - question
Matthew Dillon
dillon at apollo.backplane.com
Wed Nov 28 10:05:23 PST 2012
In the past we've found that many of these so-called benchmarks
are so poorly written that they don't actually test what they say they
are testing. For example, quite a few of them wind up doing malloc()
calls in critical loops, or gettimeofday(), or other unnecessary
system calls, and stupid things like that. And as Alex said,
a large chunk of any cpu benchmark that isn't written directly in
assembly is going to test the compiler more than it will test the
operating system.
Similarly, file I/O benchmarks often focus on only reading or only
writing and don't reflect the reality of mixed loads that most real-world
systems have to contend with.
Network benchmarks often test single-threaded or single-stream
performance, which is pretty much worthless, far more than they
test concurrent stream performance and fairness which servers are
more likely to have to deal with.
Benchmarks sometimes can identify bottlenecks and other issues that are
worthy of action. The recent postgres/pgbench tests identified some
significant issues that we were able to address in the release, for
example.
I glanced at that posting a week or three ago and generally speaking
the more recent DragonFly did do marginally better, probably due to
the positive effects the scheduler changes have on the cpu caches.
-Matt
Matthew Dillon
<dillon at backplane.com>
More information about the Users
mailing list