p7 vs p8 Real World Core Reductions

Posted by on 16 Jul 2017 in Power System Affinity, AIX Tuning, Performance Testing

We recently ran some new release load testing using Load Runner and the same workload was applied to two different LPARs with the exact same configuration. The tests were conducted on p7 and p8 hardware.

These tests show the reduced core requirements for a real world p7 to p8 migration.

LPAR Details - All LPARs identical.

AIX: 7100-04-03-1642

Shared-SMT (smtctl -t 2)

schedo vpm_throughput_mode = 2

Entitlement = 3.0

vCPU Count = 30

Memory = 40960 Megabytes - Dedicated.

p7 Details.

9119-FHB ==> p795

Clock Speed: 4004 MHz

All 256 Cores are activated.

Processor Implementation Mode: POWER 7

Firmware Version: AH780_068

p8 Details.

9119-MHE ==> E880

Clock Speed: 4024 MHz

Only 148 of the 192 Cores were activated.

Processor Implementation Mode: POWER 7

Firmware Version: FW860.20 (SC860_082)

Test Summary.

The testing injects over 1,400 transactions per second to the LPARs using load running. This ensures that the same load is allocated to each LPAR.

Test on p7 vs p8.

LPAR1 was on Node 5 in Frame 1 on p7.

LPAR2 was on Node 7 in Frame 1 on p7.

LPAR1 was on Node 3 in Frame 2 on p8.

LPAR2 was on Node 2 in Frame 2 on p8.

Shared Processor Pool Consumption.

LPAR1 around 17 to 18 cores on p7.

LPAR2 around 18 to 19 cores on p7.

LPAR1 around 14 cores on p8.

LPAR2 around 15 cores on p8.

The same workload on p8 requires around 4 less cores from the shared pool.

LPAR1 p7 Cores to User and System

p7 795 LPAR Cores Consumed

LPAR2 p7 Cores to User and System

p7 795 LPAR Cores Consumed

LPAR1 p8 Cores to User and System

p8 E880 Cores Consumed

LPAR2 p8 Cores to User and System

p8 E880 LPAR Cores Consumed

Virtual Processor Context Switching.

ilcs is the number of involuntary logical processor context switches. VP was kicked off the core by the Hypervisor due to another LPAR running under entitlement needing to run on the core.

vlcs is the number of voluntary logical processor context switches. VP was voluntary ceeded back to the shared pool as the LPAR did not have any active workload to run on the core.

LPAR1 shows around 5,000 ilcs and 14,600 vlcs on p7.

LPAR2 shows around 6,000 ilcs and 15,700 vlcs on p7.

LPAR1 shows around 4,900 ilcs and 13,300 vlcs on p8.

LPAR2 shows around 5,000 ilcs and 15,000 vlcs on p8.

LPAR1 ilcs to vlcs on p7

p7 795 ilcs vs vlcs good affinity