[Mono-list] lies and microbenmarks (Was: Mono 0.23 windows installer)

Paolo Molaro lupus@ximian.com
Tue, 11 Mar 2003 12:32:34 +0100


On 03/09/03 David Jeske wrote:
>   - CSC speeds up MS more than Mono (somewhat expected)

mcs inserts useless additional variables with the pre/post
increment/decrement operators, and that makes the jit produce somewhat
worse code, there is already a bug filed for that.

>   - while many of the mono slowdowns seem related to JIT
>     optimizations (which the new JIT will change and improve),
>     the 2x slowdown in index_string_test points to a string,
>     hashing, memory allocation, or other hashtable inefficiency

Yes, this issue should be investigated. We already discussed in the past
about storing the calculated hashvalue for strings: that may help
somewhat. Serge measured our hashtable implementation using the MS
runtime and the speed was comparable to their implementation.
The other issue is memory allocation: in the tests a lot of boxing
operations take place, I have a simple patch that speeds up both
index_test and index_string_test by about 10%, though it can be improved.

lupus

-- 
-----------------------------------------------------------------
lupus@debian.org                                     debian/rules
lupus@ximian.com                             Monkeys do it better