Default value for OPENBLAS_THREAD_LIMIT

["Dr. Thomas Orgis" <thomas.orgis%uni-hamburg.de@localhost>]

Hi,

I am in the process of adding BLAS variants with 64 bit indexing (this
is what octave and probably R actually want on a 64 bit platform)

I noticed an open question in the current state of math/openblas*:

# Default checks number of cores of build host.
# Do we want a pkgsrc-wide default?
OPENBLAS_THREAD_LIMIT?=	# empty
.if !empty(OPENBLAS_THREAD_LIMIT)
MAKE_FLAGS+=		NUM_THREADS=${OPENBLAS_THREAD_LIMIT:Q}
.endif
BENCHMARK_MAXTHREADS?=	8

What default value should be use for OPENBLAS_THREAD_LIMIT? A number
that depends on the build host is basically random, so not good.
The benchmark thread count is less of an issue, it is only used for the
benchmarking target.

I see Debian using 256, which should suffice for some time. A single
host has between 32 to 128 cores easily nowadays, in certain price
regions. Even 256 is actually no limit with 4 sockets …

I do not now how bad the overhead is with big numbers. It is also a
fact that the benefit of ever larger thread counts in NUMA setups is
not obvious.

Give me _some_ number. Anything from 8 to 512 might make sense.
Opinions?


Alrighty then,

Thomas

-- 
Dr. Thomas Orgis
HPC @ Universität Hamburg