Originally posted by: Karl-HeinzBraun

Hi all gurus and AIXperts,

VP folding is a nice performance feature. The new vpm_throughput_mode tunable even enforces VP folding.

However, VP folding has a cost in a multi-SRAD environment. When the LPAR is folded down to one single VCPU and if this LPAR is spread over several SRADs, memory access can become "far" most of the time. ASO does not help in this case bacause you do not want all memory to be moved locally to that single remaining core.

My question is: is there an option to limit the VP folding so it keeps at least one VP unfolded per SRAD ?

In this example of a badly fragmented 13-VP LPAR, I would expect at least 1 VP per SRAD (i.e 5 in total) being unfolded. This would help ASO a lot!

$ lssrad -av
REF1   SRAD        MEM      CPU
0
          0   36422.44      0-11 16-19 28-31 48-51
          2   14691.00      20-23 36-39
1
          1   18426.00      12-15 24-27 44-47
          3    9213.00      32-35
2
          4    5460.44      40-43
 

What is your opinion ? What needs to be done, so that IBM implements this feature if it does not exist yet ?

Best regards,

KH

BTW, there is a nasty problem with ASO and multi-SRAD under 6.1 TL9 and 7.1 TL3 (see APAR IV46607 and IV46579 respectively). With the option proposed here, this system crash could never occur.