API for multi-thread apps

T(DesignDocument)?

Why write a multi-threaded app?

The average number of cores per PC will increase over the next few years, possibly at a faster rate than the average amount of available RAM.

Depending on your application and project, it may be desirable to develop a multi-threaded application. Possible reasons to do this:

• If your application's memory footprint is large enough that, on some PCs, there's not enough RAM to run a separate copy of the app on each CPU.

• If you want to reduce the turnaround time of your jobs (either because of human factors, or to reduce server occupancy).

Writing and debugging a multi-threaded app is often hard. You may be able to use existing libraries of numerical "kernels" that are already multi-threaded.

Assumptions

A 'multi-thread app' A uses multiple threads, say Nthreads(A). The average number of processors used, Ncpus(A), may be less (because of I/O or synchronization).

Ideally, on a host with N CPUs, we want Ncpus(A), summed over running apps, to be about N. If it's less, we're not using CPU time. If it's more:

• we increase latency without increasing throughput
• we use more RAM than needed
• higher synchronization overhead

We assume that applications may be able to change Nthreads(A) dynamically in response to hints from BOINC. Nthreads(A) need not be equal to the hint.

Example: suppose

• we have an 80-core CPU
• app A can use 1,2,4,8,16,32 threads
• app B can use 1,2,4,8,16,32,64 threads

Then we want to have either (16,64) or (32,32) threads most of the time.

Proposal

API functions:

int boinc_target_nthreads();
void boinc_actual_nthreads(int);

An application calls boinc_target_nthreads() periodically, at points where it is able to change its number of threads. It calls boinc_actual_nthreads() to report its actual number of threads.

A WU DB record can specify "max average ncpus", an estimate of Ncpus(A) on a host with arbitrarily many CPUs. This is used by the client and scheduler to estimate completion time.

Implementation

Shared-memory messages:

• core->app (process control channel): <target_nthreads>
• app->core (process control channel): <actual_nthreads>

Client maintains estimates of CPU efficiency per job, uses this to scale target_nthreads.

Implementation (enforce_schedule()): as we schedule jobs, decrement CPU count by scaled actual_nthreads. rr_simulation() needs to be modified too.