|
DisCostiC v2.0.0
Distributed Cost in Cluster: a cross-architecture simulator for massively parallel MPI programs
|
Part of the main README.
DisCostiC selects behavior with several independent modes. Collective mode is one of them. Kernel modes choose how a compute kernel is timed. Multi-kernel mode chains two compute phases. Heterogeneous mode places ranks on two system descriptions. exchange_mode (message size, step, direction, periodicity) stays in the configuration tables of the main README.
kernel_mode is COMP, LBL, FILE, or SRC. The summary table is in the main README. FILE and SRC call Kerncraft. COMP and LBL do not. The Exec string carries the mode:
Headers follow <BENCHMARK>_<MODE>.hpp under test, for example STREAM_FILE, STREAM_SRC, STREAM_LBL, and STREAM_COMP. Set benchmark_kernel in config.cfg to that header name without .hpp.
Set collective_mode = 1 in config.cfg. The simulator then runs one timestep. heteregeneous must be 0. Combining collective_mode = 1 with heteregeneous = 1 is rejected at startup.
test/STREAM_COLLECTIVE.hpp runs a STREAM Exec and then DisCosTiC->Collective(...). The header reads three keys from config.cfg. The shipped config.cfg does not define them, so a STREAM_COLLECTIVE run needs these lines added:
| Key | Role |
|---|---|
collective_type | DisCosTiC::CollectiveType: COLL_BCAST = 0, COLL_REDUCE = 1, COLL_ALLREDUCE = 2 |
collective_algorithm | Algorithm id for that type |
collective_bytes | Collective payload in bytes |
| Type | Algorithm enum | Value |
|---|---|---|
COLL_BCAST | BCAST_BINOMIAL | 0 |
COLL_BCAST | BCAST_RING | 1 |
COLL_REDUCE | REDUCE_BINOMIAL | 0 |
COLL_ALLREDUCE | ALLREDUCE_RECURSIVE_DOUBLING | 0 |
include/CollectiveModel.hpp builds the broadcast, reduce, and allreduce schedules from LogGP or Hockney parameters for intra-chip, inter-chip, and inter-node scopes. The call shape used by STREAM_COLLECTIVE is:
Two compute phases in one timestep use part1 and part2 in config.cfg. Each phase has its own Exec mode (FILE, LBL, SRC, or COMP). The shipped config sets part1 = STREAM and part2 = TRIAD.
| Header | What it chains |
|---|---|
TWOKERNEL | Two compute kernels selected by part1 / part2 |
PITRIADFILE | Pi and STREAM/TRIAD in FILE mode |
PITRIADLBL | Pi and STREAM/TRIAD in LBL mode |
RTOF4 | Ray tracing and optical flow in LBL mode |
Kerncraft and the per-rank ECM config copy run once per kernel, not once per modeled rank. See include/AST.hpp.
heteregeneous is spelled that way in config.cfg and in src/DisCosTiC.cpp.
heteregeneous | System |
|---|---|
0 | One system: number_of_processes ranks on micro_architecture |
1 | Two systems: primary ranks plus secondary_number_of_processes ranks on secondary_micro_architecture |
Secondary keys in config.cfg include secondary_task_per_node, secondary_compiler_flags, secondary_pmodel, secondary_N, secondary_dim_x, secondary_dim_y, and secondary_penalty. This mode is independent of collective mode and cannot be combined with collective_mode = 1.