Can add compute templates which can be used to deploy images.
- home node is core that a VP first gets dispatched to
- will try to use home node where possible
- p5/6 loads 1st & 2nd SMT to ~80% before unfolding next VP
- p7 loads first thread to 50% then unfolds next VP
- when all VPs unfolded, starts using 2nd SMT thread
- when all 2nd threads loaded, starts using tertiaries
- provides highest throughput at expense of more cores
- both report same physical usage
- more cores using in p7
- idle time higher
- reduce VPs accordingly
- behaves more like P6 - not recommended; try to reduce VPs instead
schedo -p -o vpm_throughput_mode=blah
- SMT threads go with VPs
- p5/6 SMT thread can go to 100%
- p7 SMT thread can only go up to 63%
- if only primary SMT threads are being used, you have too many VPs
- If ent not long enough, VPs spend a lot of time warming cache
- put this somewhere between peak and average
- don't set num VPs higher than cores in processor pool
lparstat
, pool size, available in pool, virt context switches, phantom interrupts (vm dispatched to core, then ceeded on IO, but after IO interrupt comes back when target VM not running)sar -mu -PALL
vertically, all threads with, physc & %entcmpstat -s
shows procs horizontallyvmstat -IW 60 2
columns:p
threads waiting for IO to raww
column is how many threads waiting for CIO/DIO.
- %comp > 97% means you page
- comp is also called dynamic or working storage
lssrad -av
shows local/near memory for cpus
- try to avoid paging in first place
- use multiple, all same size on different disks
lru_file_repage=0
(restricted in AIX7) means care only about minperm- on AIX 5.3TL4, should set maxperm & maxperm to 90, only change them with PMR
page_steal_method
default is0
.1
means scan only filesystem pages, reduces scanning time. When combined with CIO can make a big difference.
vmstat -v
don't want to see high IO blocked
numperm
= numclient + jfs and other stuffnumclient
jfs2 & network- if
numperm
==numclient
then it is all being used by network or jfs2 - pbuf is where IO initially gets built
vmstat -v | grep memory
find number of memory pools- minfree & maxfree apply to memory pools
- vmstat fre column needs to be divided amongst memory pools