Performance

SMT8 on IBM Power

What SMT8 is, the real rPerf spread between single-thread and SMT8 on Power 11 hardware, and when eight threads per core help a workload rather than just inflating a benchmark.

SMT8 is simultaneous multithreading with eight hardware threads per physical core, the deepest threading mode on IBM Power. It is also the mode most Power benchmark figures are quoted at, which is why a Power 11 rPerf number can look implausible next to the core count. The gap between single-thread and SMT8 on the same silicon is roughly threefold, and whether your workload sees that gap or not is the entire question.

Threads per core8 at SMT8
Modes availableST, SMT2, SMT4, SMT8
IntroducedPOWER8
Typical ST to SMT8 gainAround 3x on published rPerf
Set perPartition, not per system
Licensing effectNone. Licensing follows cores, not threads

The real spread, from IBM's own published figures

These are the published rPerf ratings for actual Power 11 processor features at each threading level. The same physical hardware, measured four ways.

rPerf by SMT level on published Power 11 features

ModelFeatureCoresSTSMT2SMT4SMT8
S1112EJMT439.981.8119.9148.8
S1112EJSV1095.4185.6247298.7
S1122ERGR879.9159.8221.4292.9
S1122ERGQ20173.2338.1482.8626.7
S1122EBG832294.2567.2806.91,040.3
S1122EBGA60488.2913.41,191.81,531.9
S1124EP3X32305.8629.6891.91,117.1
S1124EP3Y60515.81,0521,374.21,736.8

Note the shape of the curve. The jump from ST to SMT2 is close to double every time. From SMT4 to SMT8 the gain is far smaller, typically 25 to 30 percent. Threads are not free capacity, they are better utilization of an idle pipeline.

When SMT8 earns its place, and when it does not

Good fit: many concurrent transactions

Classic IBM i commercial workloads, interactive users, web and API serving, and anything with a large number of independent units of work. These keep the pipeline fed and see close to the published SMT8 numbers.

Poor fit: a single long batch job

A nightly batch that cannot be parallelized runs on one thread and sees the ST figure, not the SMT8 one. If your month-end window is the constraint, size on ST rPerf and pick a high-clock feature code, not a high core count.

Cache-bound work

Eight threads share the core's 2 MB L2 and 8 MB L3 on Power 11. A workload with a large working set can thrash cache at SMT8 and run measurably better at SMT4. This is worth testing rather than assuming.

Licensing is unaffected

SMT mode does not change what you pay. IBM i software tiers and per-core licensing follow activated cores, so raising or lowering SMT costs nothing and risks nothing commercially. It is a free experiment.

Because SMT is set per partition, a single system can run one LPAR at SMT8 for transaction work and another at SMT4 or ST for a batch partition. That is usually a better answer than picking one mode for the whole machine.

How to check and change the SMT mode

On IBM i

The QPRCMLTTSK system value controls processor multitasking. Setting 1 enables it, 0 disables it, and 2 lets the system decide. Combine with WRKSYSACT to see what the change actually does to your workload rather than trusting the benchmark.

On AIX

smtctl reports the current mode and thread count. smtctl -m off drops to single-thread, and smtctl -t 4 or -t 8 sets the thread count explicitly. Changes can be made dynamically or persisted across reboot with -w boot.

On Linux on Power

ppc64_cpu --smt reports the current mode and ppc64_cpu --smt=4 or --smt=8 sets it. lscpu confirms the resulting thread-per-core count.

Measure, do not assume

Run the real workload at SMT4 and SMT8 and compare. The published figures are a guide to the shape of the curve, not a prediction for your application. The test is cheap and reversible.

Related

Every Power 11 processor feature and its benchmark figures are on the Power 11 CPU page. For what rPerf and CPW actually measure, see rPerf and CPW, and the short definition of this term is at SMT8 in the glossary. Per-model configuration ceilings are on the Power 11 specs page.