cpu/vcpu/lcpu ?

[tempest92]

can someone please confirm or comment on my understanding of cpu/core/vcpu/lcpu is correct -

profile -
uncapped partiton
proc unit 0.1/0.4/8
vcpu 1/4/8

so each vcpu will have a guaranteed 10% of a processing unit (a core ? - 16 cores on an 8 cpu dual core server ?)

as the partition is uncapped, should the partition need and the shared processor pool have spare cycles,
each vcpu could go up to 100% of a processing unit.

with SMT enabled, each vcpu appears as 2 lcpu's but the boundary is obviously still 100% of 4 vcpus
(processing units)

Thanks

 

[sjm2]

so each vcpu will have a guaranteed 10% of a processing unit
Simon says: yes

(a core ?
Simon says: yes

- 16 cores on an 8 cpu dual core server ?)
Simon says: depends what you mean by a cpu. IBM doesn't really use the term "dual core server" with POWER machines. This is because when IBM started making "chips" with two "execution units" per chip, since IBM predominantly makes servers, it makes sense to say that we have a server with a certain number of processors. Intel predominantly make chips, which they call processors. So if a "chip" has two "cores" it makes sense for them to call it a dual core processor. Thus, a 16 core System p 570 has eight POWER5 chips, each with 2 cores. In IBM land that's 16 processors - in Intel land, that's 8 processors.

as the partition is uncapped, should the partition need and the shared processor pool have spare cycles,
each vcpu could go up to 100% of a processing unit.
Simon says: this rather depends. Rather say, since it is uncapped, it can get the lower of (a) the processing units in the pool and (b) the number of vcpus*1.0 that the SPLPAR currently has. See "Simon adds".

with SMT enabled, each vcpu appears as 2 lcpu's but the boundary is obviously still 100% of 4 vcpus
(processing units)
Simon says: let us not say "boundary" but maximum. An SPLPAR cannot consume more than vcpus*1.0, as there are no vcpus to adsorb the cycles.

Simon adds: Remember that these "triples" are values for minimum/desired/maximum. So, looking at an extreme case and remembering that the minimum of 0.1 processing units per vcpu is a hard lower limit, if there were only the minimum of 0.1 processing units left uncommitted after starting other SPLPARs, then this SPLPAR could still be started but would only be started with ONE vcpu, as starting it with two would violate the minimum 0.1 processing unit per vcpu rule. Thus, your SPLPAR may be started with only one vcpu. In which case (and without PLM taking action), even if there were cycles available, this partition could only use 100% of one vcpu.