Conversation
Petr Tesarik
ptesarik@fosstodon.org
1
0
3
1
0
2
@hyeyoo @ptesarik @vbabka it's how those ranges are mapped/made available that is the issue and also the idea of networked memory starts to get crazy pretty quick.
Just a case of doing things sensibly, feels like there's 2 camps for CxL - the sensible, slot it in in a sane way camp and the GFP_LET'S_GOOOOOO camp
Just a case of doing things sensibly, feels like there's 2 camps for CxL - the sensible, slot it in in a sane way camp and the GFP_LET'S_GOOOOOO camp
1
0
0
@ljs @ptesarik @vbabka
CXL is not feasible for unmovable kernel allocations regardless of latency if the goal is reduction in total memory cost (and thus dynamic capacity change mechanism is used)
BTW the latency is not that bad, it's just like a far and far NUMA node...
ofc emulated CXL is a complete beast - someone recently reported having 500KB/s bandwidth with it ;)
CXL is not feasible for unmovable kernel allocations regardless of latency if the goal is reduction in total memory cost (and thus dynamic capacity change mechanism is used)
BTW the latency is not that bad, it's just like a far and far NUMA node...
ofc emulated CXL is a complete beast - someone recently reported having 500KB/s bandwidth with it ;)
1
0
0
@hyeyoo @ptesarik @vbabka yeah I think the sensible way of implementing it is as a CPU-less node. That's fine.
It's when you're saying 'no we need a new mmap() flag and a new gfp flag and we need to change all the code to suit it' that there's a problem.
Reminds me of another proposal, but my lips are SEALED about which one.
500KB/s well I guess you don't want to test anything too crazy then :P
It's when you're saying 'no we need a new mmap() flag and a new gfp flag and we need to change all the code to suit it' that there's a problem.
Reminds me of another proposal, but my lips are SEALED about which one.
500KB/s well I guess you don't want to test anything too crazy then :P
1
0
2
Petr Tesarik
ptesarik@fosstodon.org
0
0
3