argonne-lcf / alcf-mpi-benchmarks Goto Github PK

ALCF MPI benchmarks suite consists of five independent programs: mmps, pingpong,
aggregate, bisection, and collectives, which measure and report,
correspondingly, a zero length messaging rate, point-to-point communication
latency with ping-pong benchmark, single node aggregate communication
bandwidth, bisection bandwidth, and the latency of certain collective
operations.

Messaging rate “mmps”: To maximize the rate, the offeror may choose appropriate
number of communicating neighbors XNBORS, the number of MPI tasks placed on the
reference node, and the window size - the parameter, which determines the
number of messaging sent simultaneously to each neighbor. The placement of the
neighbors is defined by proper modification of the “getranks.c” file according
to specifics of the interconnect subsystem. 

Point-to-point communication latency is measured by a “pingpong” benchmark for
three cases: intranode, nearest neighbor, and farthest path. The offeror can
modify the “getranks.c” file to specify the communicating tasks according to
specifics of the interconnect. For intranode benchmark, the communicating tasks
must reside on the same node. For nearest neighbor benchmark, the communicating
tasks must be placed on two distinct nodes, connected by the nearest path,
provided by the interconnect subsystem. For the farthest path benchmark, the
two communicating tasks must be placed on two distinct nodes with the longest
interconnect path between them.

Single node aggregate bandwidth benchmark  measures and reports the total 
aggregate interconnect bandwidth, available to a task by aggregating  the
bandwidth of all available links on the node. The benchmark utilizes
point-to-point communications. To maximize the bandwidth, the offeror may
specify “N” - the number of communicating tasks. The offeror may also modify
the “getranks.c” file to place the communicating tasks according to specifics
of their proposed interconnect subsystem.

Bi-section bandwidth benchmark is measures and reports the aggregate bandwidth
between the worse-case bi-section of the partition. The offeror should generate
the bi-section by modifying the “getranks.c” file according to specifics of
their interconnect subsystem.

The latency of collective operations, specifically Barrier, Broadcast, and
Allreduce, is measured and reported by a “collective” benchmark. The
measurements are performed across entire partition via World communicator, as
well as the two non-overlapping sub partitions, dividing the World in two equal
parts. The offeror may specify the division of the World into parts by
modifying the “split.c” file, according to their preference and using specifics
of the interconnect subsystems.