Linux NUMA

1
(No Transcript)
2
Linux NUMA
John Wright Engineering Manager Core Linux
3
NUMA plans

• First pass NUMA support
• DISCONTIGMEM support (already in 2.4)
• First touch default allocation policy (already in
  2.4)
• Node specific page allocation (already in 2.4)
• Node specific bootmem allocation (already in 2.4)
• Kernel text replication (arch specific, in 2.4
  for MIPS64 port)

4
NUMA plans

• First pass NUMA support
• Per node kmalloc
• Per node kswapd/page out
• runon / cpuset functionality (mechanism in 2.4,
  API still being worked on)
• Heirarchical scheduling policy (probably generic
  calls to arch specific cpu affinity and node
  affinity algorithms)
• Basic dplace functionality/memory placement API

5
NUMA plans

• Second pass NUMA support
• More advanced dplace functionality (such as
  dprof)
• More advanced cpuset functionality (hard and soft
  affines), advanced scheduler options
• Multipath i/o with NUMA awareness (dependent on
  SCSI infrastructure development, XSCSI has latent
  multipath features)
• NUMAize obvious kernel structures via
  benchmarking data (possibly need queued
  spin-locks to avoid NUMA caused starvations)

6
NUMA plans

• Futures
• Explore alternative memory allocation policies
• User text replications (possible little benefit)
• Memory migration, process migration (complicated
  and possible little benefit depending on link
  bandwidth)
• More NUMAization of structures based on more
  benchmarking data and generic linux kernel
  evolution (will always be ongoing work)

7
(No Transcript)