On 02/22/2013 04:41 PM, Rik van Riel wrote:
On 02/22/2013 10:35 AM, Larry Woodman wrote:
1.) Using mmu notifiers to set up the page tables for integrated
devices(GPUs) and allowing the generic
kernel pagefault handler to resolve translations for those
devices.
This functionality is also desired by people who want to run
memory coherency over infiniband and some other very fast
network fabrics, as well as by the people who want to offload
work to specialized cores in their system.
Since there are multiple use cases for this kind of functionality,
I believe that it is important we get the infrastructure right,
and discussing this topic at LSF/MM sounds like a good idea.
Agreed.
2.) Replication of pagecache pages on NUMA nodes.
What about this would you like to discuss?
The tradeoffs between local memory access for frequently referenced
read-only/executable
pages like shared library text and the additional memory consumption
associated replicating
pages on multiple NUMA nodes.
Is there some proposal of code to do this?
Not yet, I implemented replication of filesystem cache memory on each
NUMA node
several years ago for a UNIX kernel when NUMA systems first appeared.
At that
time both memory and caches were much smaller than they are now so the
tradeoffs
between local memory access and inducing page reclamation were different
than they
are now. However there was a significant performance boost on memory
rich systems
running applications with bloated text sections.
Larry
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>