|
| 1 | +LPGROUPS |
| 2 | +{ |
| 3 | + MODELNET_GRP |
| 4 | + { |
| 5 | + repetitions="36"; |
| 6 | +# name of this lp changes according to the model |
| 7 | + nw-lp="2"; |
| 8 | +# these lp names will be the same for dragonfly-custom model |
| 9 | + modelnet_dragonfly_dally="2"; |
| 10 | + modelnet_dragonfly_dally_router="1"; |
| 11 | + } |
| 12 | +} |
| 13 | +PARAMS |
| 14 | +{ |
| 15 | +# packet size in the network |
| 16 | + packet_size="4096"; |
| 17 | + modelnet_order=( "dragonfly_dally","dragonfly_dally_router" ); |
| 18 | + # scheduler options |
| 19 | + modelnet_scheduler="fcfs"; |
| 20 | +# chunk size in the network (when chunk size = packet size, packets will not be |
| 21 | +# divided into chunks) |
| 22 | + chunk_size="${CHUNK_SIZE}"; |
| 23 | +# modelnet_scheduler="round-robin"; |
| 24 | +# number of routers in group |
| 25 | + num_routers="4"; |
| 26 | +# number of groups in the network |
| 27 | + num_groups="9"; |
| 28 | +# buffer size in bytes for local virtual channels |
| 29 | + local_vc_size="16384"; |
| 30 | +#buffer size in bytes for global virtual channels |
| 31 | + global_vc_size="16384"; |
| 32 | +#buffer size in bytes for compute node virtual channels |
| 33 | + cn_vc_size="32768"; |
| 34 | +#bandwidth in GiB/s for local channels |
| 35 | + local_bandwidth="5.25"; |
| 36 | +# bandwidth in GiB/s for global channels |
| 37 | + global_bandwidth="4.7"; |
| 38 | +# bandwidth in GiB/s for compute node-router channels |
| 39 | + cn_bandwidth="5.25"; |
| 40 | +# ROSS message size |
| 41 | + message_size="840"; |
| 42 | +# number of compute nodes connected to router, dictated by dragonfly config |
| 43 | +# file |
| 44 | + num_cns_per_router="2"; |
| 45 | +# number of global channels per router |
| 46 | + num_global_channels="2"; |
| 47 | +# network config file for intra-group connections |
| 48 | + intra-group-connections="@CMAKE_SOURCE_DIR@/src/network-workloads/conf/dragonfly-dally/dfdally-72-intra"; |
| 49 | +# network config file for inter-group connections |
| 50 | + inter-group-connections="@CMAKE_SOURCE_DIR@/src/network-workloads/conf/dragonfly-dally/dfdally-72-inter"; |
| 51 | +# routing protocol to be used |
| 52 | + routing="prog-adaptive"; |
| 53 | + minimal-bias="1"; |
| 54 | + df-dally-vc = "1"; |
| 55 | +# counting msgs recv to/send from routers |
| 56 | + counting_bool="0"; |
| 57 | + counting_start="0"; |
| 58 | + counting_windows="1800"; |
| 59 | + #interval in us |
| 60 | + counting_interval="300"; |
| 61 | + num_apps="2"; |
| 62 | + #offset for app_id: model-net-mpi-replay is 88, synthetic-dfly-plus is 24 |
| 63 | + offset="144"; |
| 64 | +} |
| 65 | +NETWORK_SURROGATE { |
| 66 | + enable="${NETWORK_SURR_ON}"; # either "0" or "1" |
| 67 | +# determines the director switching from surrogate to high-def simulation strategy |
| 68 | + director_mode="delegate-to-app-director"; |
| 69 | + #director_mode="at-fixed-virtual-times"; |
| 70 | + |
| 71 | +# director configuration for: director_mode == "at-fixed-virtual-times" |
| 72 | +# timestamps at which to switch to surrogate-mode and back |
| 73 | + fixed_switch_timestamps=( "25.0e6", "400.0e6" ); |
| 74 | + |
| 75 | +# latency predictor to use. Options: average, torch-jit |
| 76 | + packet_latency_predictor="average"; |
| 77 | +# some workload models need some time to stabilize, a point where the network behaviour stabilizes. The predictor will ignore all packet latencies that arrive during this period |
| 78 | + ignore_until="2.0e6"; |
| 79 | + |
| 80 | +# parameters for torch-jit latency predictor |
| 81 | + torch_jit_mode="single-static-model-for-all-terminals"; |
| 82 | + torch_jit_model_path=""; |
| 83 | + |
| 84 | +# selecting network treatment on switching to surrogate. Options: freeze, nothing |
| 85 | + network_treatment_on_switch="${NETWORK_MODE}"; |
| 86 | +} |
| 87 | +APPLICATION_SURROGATE { |
| 88 | + enable="${APP_SURR_ON}"; # either 0 or 1 |
| 89 | + |
| 90 | + # Configuring director |
| 91 | + director_mode="${APP_DIRECTOR_MODE}"; # Opts: "every-n-gvt", "every-n-nanoseconds" |
| 92 | + director_num_gvt="${EVERY_N_GVT}"; |
| 93 | + director_num_ns="${EVERY_NSECS}"; # 1^6 means 1ms |
| 94 | + |
| 95 | + # Configuring predictor |
| 96 | + # Minimum number of iterations to collect data from before skipping ahead in the simulation |
| 97 | + num_iters_to_collect="2"; |
| 98 | +} |
0 commit comments