[libfabric-users] Cache Coherency

[david Atkins]

We are using libfabric to move uncompressed video between compute nodes in  a cluster, I'm sure my concern is handled within the libraries but I'm wondering how the CPU caches are invalidated such that if we write between nodes that the receiver knows the memory needs to be re-read into cache.

I ask because we are looking at an issue and while I'm sure this is all correctly handled but  I am intrigued as to the mechanism that is used to ensure that all local caches are invalidated when a buffer is submitted for use as a receive target.

There is a risk we are re-queuing a buffer to soon and we need to guard against this.



David Atkins

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/libfabric-users/attachments/20240301/940af2b2/attachment.htm>