Red Hat Bugzilla – Bug 257301
RHEL 5 is 20% slower than RHEL 4.x on same hardware
Last modified: 2007-11-30 17:07:47 EST
Description of problem:
On the identical hardware, RHEL 5 Xen-enabled kernel is 20% or more slower than
Version-Release number of selected component (if applicable):
Consistently. On HP dx5150 or Dell Precision 390.
Steps to Reproduce:
1. Use two identical machines or load RHEL4 run tests first and then reload
with RHEL 5 and run the same tests as below
2. Untar kernel-188.8.131.52 to a scratch directory on local disk
3. Run "cd /tmp/linux-184.108.40.206; make oldconfig", then run "time make".
RHEL5 takes 20% or longer to finish compiling the same kernel than RHEL4.
RHEL5 takes same amount of time or less to compile linux kernel as RHEL4.
Here are some test results:
$ time tar -zxvf linux-220.127.116.11.tar.gz
$ time make
You probably wanted to select a component "kernel".
So, RHEL-5 is actually something like 50% slower here.
But... this does not look like a kernel issue. It's certainly not a Xen issue,
as the difference between the RHEL-5 xen and non-xen kernels is tiny compared to
the difference between RHEL-4 and RHEL-5.
But RHEL-5 has a newer compiler that is doing more, and different,
optimisations; it is entirely reasonable that it may take longer for some jobs.
Indeed, the numbers above show that the bulk of the difference is accounted for
in user time, ie. time spent in the compiler, not the kernel.
We cannot consider this to be a kernel issue unless it is the same compiler
version being tested in each case.
RHEL4 has gcc 3.4, which had only RTL optimization passes, gcc 4.1 has both
tree-SSA and RTL optimization passes, so roughly 100 new optimization passes.
While a lot of work has been done to improve compile speed and it is still
ongoing (e.g. GCC 4.3 will have some compile time speed improvements), some
slowdown from 3.4.x is to be expected. You get for that price better optimized
code in many cases and also some new GCC extensions that aren't really doable
on the RTL representation (e.g. -D_FORTIFY_SOURCE with gcc 3.4.x is orders
of magnitude worse protection than in gcc 4.0.x+).