[ltp] Re: Memory bandwith result from Memtest 86+?

Daniel Pittman linux-thinkpad@linux-thinkpad.org
Fri, 30 Nov 2007 11:41:40 +1100


jqian@physics.harvard.edu writes:

>> I wouldn't trust memtest86+ as an effective benchmark, not least of
>> which because it really isn't designed to benchmark so much as stress
>> test.
>
> What is a good benchmark for memory bandwidth under linux?

No idea, sorry. :/

>> Sadly, no.  The biggest thing that will notice the main memory speed
>> is every little thing you do, especially interactive GUI tools,
>> because they need more data than fits into your cache.
>> 
>> Unless you run very CPU intensive processes (like a compiler) that
>> need a limited quantity of data (for small chunks of code) then you
>> usually fall out of cache into slower memory.
>> 
>> Your web browser, for example, will comfortable use more than the 8MB of
>> L3 cache you get in the best equipped CPU these days -- let alone what
>> shows up in the Thinkpad mobile CPU lines. 
>
> Unless you get one of these monsters
> http://en.wikipedia.org/wiki/Power6
> 32MiB of off die L3 Cache! I suppose firefox, fixing its memory leak, 
> can be made fitting into those.

If you find a thinkpad with a Power6 in it I will happily buy one. ;)

>> Faster memory is the second /best/ investment you can make to improve
>> overall performance of your laptop, following only faster disks.
> Unfortunately the upper limit of memory is set by the bus speed of a 
> laptop, which is not easy to change for an laptop. For example, my 
> laptop have a bus speed of 533MHz of its mainboard.
>
> My disk is 7200RPM disk, which hdparm /dev/sda:
>  Timing cached reads:   3436 MB in  2.00 seconds = 1719.02 MB/sec
>  Timing buffered disk reads:  114 MB in  3.02 seconds =  37.72 MB/sec
>
> Is that fast for a laptop? What's is that cached read? Is that a good 
> benchmark for memory bandwidth?

Yes, that is very fast.  The "cached read" bit is reading from the Linux
buffer cache (main memory, pretty much), while the second is reading
from the cache memory on the hard disk.

> Thanks for the education.

Happy to help.  Sadly, understanding performance on modern systems is
very, very difficult -- much more than it was ten years ago when it was
hard to learn enough.

(Oh, and I can't actually comment on the Power6 CPU really, as I don't
 actually know enough about their design and layout to really
 understand.)

Regards,
        Daniel
-- 
Daniel Pittman <daniel@cybersource.com.au>           Phone: 03 9621 2377
Level 4, 10 Queen St, Melbourne             Web: http://www.cyber.com.au
Cybersource: Australia's Leading Linux and Open Source Solutions Company