Juan Manuel Placco
|
6.3 vs 6.4 uK message passing comparison (same node)
|
Juan Manuel Placco
12/03/2008 1:07 PM
post17865
|
6.3 vs 6.4 uK message passing comparison (same node)
Hi,
As Robert suggest I'm moving to this forum with my question.
This cames from a large discussion in:
http://community.qnx.com/sf/go/projects.networking/discussion.technology.topc5048
Robert wrote:
>>I'm a bit surprised that the comparison shows that much
>> improvement with large message transfer sizes as well.
>> I would have expected the improvements to be most noted
>> for small messages (in which the kernel overhead dominates
>> the message pass) as opposed to the large messages
>>(which should be dominated by the actual time to do the copy).
>> I think that you might want to post this information in one of
>> the forums on the OS project to get a better explanation.
>> I'm not sure of the kernel optimizations which were introduced
>> in 6.4.0 which would affect message passing.
>> Robert.
The thing is a have a benchmark test comparing message passing (form different sizes) in the SAME node, in 6.3 and in 6.
4 installations (the same machine with 2 partitions).
I was expecting the same results as Robert mention, but I'm still surprised and wondering what's behind de scene... How
can 6.4 be much faster moving data from client to server address space? It's just something very similar to a memmove,
isn't it? And this is, as I suppose almos an assembly level function with lot of work for DMA and RAM memory.
I attach the results in an excel file.
I will be very thankful if they can help me to understand this behavior difference between 6.3 and 6.4 kernel!
Regards,
Juan Manuel
|
|
|