Re: Fastcode MM B&V 0.34

God morning

ToDo list:

- Move benchmark -> use CPUID initialized function pointer for Move calls.
- include longer Replay log from Dan Downs
http://www.retnuh.org/files/webbroker.zip
- new replay benchmark with Nathanial Woolls's replay log
- Include a version of Nexus that passes validation?
- Integrate the dll validation into the B&V under the validate button. Not
possible?!

Have some way for the replays to do more than just alloc/realloc/frees,
so that cache and alignment issues become visible (if there are any).
Not sure what this could involve exactly. Currently only the 1st
and last bytes are written too.

Eric Grange Random ideas:
- when freeing a block read/write its first bytes, to mimic
what happens in a reference counting (strings, arrays),
or when freeing objects (that'll free/notify/whatever their
subobjects are).
- randomly pick an allocated block and read its whole content
- randomly pick N allocated blocks and read their content (N>2),
from start to end in parallel (and write to 0-M random
blocks)
- read with byte/word access for allocated blocks whose requested
size wasn't a multiple of four, read with double access for
allocated blocks whose size was a multiple of 8.

For "random" we would use a deterministic random of course
(like forcing RandSeed to 0 before playback f.i.).

Another possibility could be to playback twice: once with
no other actions than alloc ones, and another time
with the read/writes added.

Regards
Dennis


.

Quantcast