makespan: add a few more plots

tasks/lammps/README.md

@@ -29,6 +29,8 @@ Upload the WASM file:
 and run the experiment with:
 
 ```bash
+# TODO: consider making this experiment shorter
+# TODO: do we need the network workload?
 (faasm-exp-faabric) inv lammps.run.wasm -w compute -w network
 ```
 

tasks/lammps/run.py

@@ -99,6 +99,9 @@ def native(ctx, w, repeats=1):
     """
     num_cpus_per_vm = 8
     num_vms = 2
+    # data_file = get_faasm_benchmark(LAMMPS_SIM_WORKLOAD)["data"][0],
+    # TODO: is this a good idea? FIXME FIXME DELETE ME
+    data_file = get_faasm_benchmark("compute")["data"][0]
 
     for workload in w:
         if workload not in LAMMPS_SIM_WORKLOAD_CONFIGS:
@@ -121,7 +124,7 @@ def native(ctx, w, repeats=1):
 
         native_cmdline = "-in {}/{}.faasm.native".format(
             LAMMPS_MIGRATION_NET_DOCKER_DIR,
-            get_faasm_benchmark(LAMMPS_SIM_WORKLOAD)["data"][0],
+            data_file,
         )
 
         for nproc in NPROCS_EXPERIMENT:

tasks/makespan/plot.py

@@ -1,5 +1,5 @@
 from invoke import task
-from matplotlib.pyplot import subplots
+from matplotlib.pyplot import subplots, subplot_mosaic
 from tasks.util.makespan import (
     MAKESPAN_PLOTS_DIR,
     do_makespan_plot,
@@ -49,3 +49,128 @@ def migration(ctx):
     # ----------
 
     save_plot(fig, MAKESPAN_PLOTS_DIR, "mpi_migration")
+
+
+@task
+def conservative(ctx):
+    """
+    Macrobenchmark plot showing the benefits of migrating MPI applications to
+    improve locality of execution. We show:
+    - LHS: box plot of idle vCPUs and # of cross-VM links for all VM sizes
+    - RHS: timeseries of one of the box plots
+    """
+    # NOTE: probably we want highter num-tasks here to make sure we migrate
+    # more
+    # num_vms = [16, 24, 32, 48, 64]
+    # num_tasks = [50, 75, 100, 150, 200]
+    num_vms = [8, 16, 24]
+    num_tasks = [25, 50, 75]
+    num_cpus_per_vm = 8
+
+    # RHS: zoom in one of the bars
+    timeseries_num_vms = 16 # 32
+    timeseries_num_tasks = 50 # 100
+
+    results = {}
+    for (n_vms, n_tasks) in zip(num_vms, num_tasks):
+        results[n_vms] = read_makespan_results(n_vms, n_tasks, num_cpus_per_vm)
+
+    fig, ax = subplot_mosaic([['upper left', 'upper right'],
+                              ['lower left', 'lower right']])
+
+    # ----------
+    # Plot 1: boxplot of idle vCPUs and num xVM links for various cluster sizes
+    # ----------
+
+    do_makespan_plot(
+        "boxplot_vcpus",
+        results,
+        ax["upper left"],
+        num_vms,
+        num_tasks
+    )
+
+    do_makespan_plot(
+        "percentage_xvm",
+        results,
+        ax["lower left"],
+        num_vms,
+        num_tasks
+    )
+
+    # ----------
+    # Plot 2: (two) timeseries of one of the cluster sizes
+    # ----------
+
+    do_makespan_plot(
+        "ts_vcpus",
+        results,
+        ax["upper right"],
+        timeseries_num_vms,
+        timeseries_num_tasks
+    )
+
+    do_makespan_plot(
+        "ts_xvm_links",
+        results,
+        ax["lower right"],
+        timeseries_num_vms,
+        timeseries_num_tasks
+    )
+
+    # ax[0][0].legend()
+    save_plot(fig, MAKESPAN_PLOTS_DIR, "resource_usage")
+
+
+@task
+def eviction(ctx):
+    """
+    Macrobenchmark plot showing the benefits of migrating MPI applications to
+    evict idle VMs.
+    - LHS: Bar plot of the VMseconds used per execution
+    - RHS: timeseries of the number of idle VMs over time
+    """
+    # NOTE: probably we want lower num-tasks here to just show the benefits
+    # at the tails
+    # num_vms = [16, 24, 32, 48, 64]
+    # num_tasks = [50, 75, 100, 150, 200]
+    num_vms = [8, 16, 24]
+    num_tasks = [25, 50, 75]
+    num_cpus_per_vm = 8
+
+    # RHS: zoom in one of the bars
+    timeseries_num_vms = 16 # 32
+    timeseries_num_tasks = 50 # 100
+
+    results = {}
+    for (n_vms, n_tasks) in zip(num_vms, num_tasks):
+        results[n_vms] = read_makespan_results(n_vms, n_tasks, num_cpus_per_vm)
+
+    fig, ax = subplot_mosaic([['left', 'right'],
+                              ['left', 'right']])
+
+    # ----------
+    # Plot 1: bar plot of the CPUsecs per execution
+    # ----------
+
+    do_makespan_plot(
+        "used_vmsecs",
+        results,
+        ax["left"],
+        num_vms,
+        num_tasks
+    )
+
+    # ----------
+    # Plot 2: timeseries of one of the cluster sizes
+    # ----------
+
+    do_makespan_plot(
+        "ts_idle_vms",
+        results,
+        ax["right"],
+        timeseries_num_vms,
+        timeseries_num_tasks
+    )
+
+    save_plot(fig, MAKESPAN_PLOTS_DIR, "idle_vms")

tasks/makespan/run.py

@@ -12,6 +12,7 @@
     EXEC_TASK_INFO_FILE_PREFIX,
     GRANNY_BASELINES,
     IDLE_CORES_FILE_PREFIX,
+    NATIVE_BASELINES,
     init_csv_file,
     get_idle_core_count_from_task_info,
     get_num_cpus_per_vm_from_trace,
@@ -73,7 +74,7 @@ def granny(
 @task()
 def native_slurm(
     ctx,
-    workload="all",
+    workload="mpi-migrate",
     num_vms=32,
     num_cpus_per_vm=8,
     num_tasks=100,
@@ -96,7 +97,7 @@ def native_slurm(
 @task()
 def native_batch(
     ctx,
-    workload="all",
+    workload="mpi-migrate",
     num_vms=32,
     num_cpus_per_vm=8,
     num_tasks=100,
@@ -152,22 +153,25 @@ def _do_run(baseline, num_vms, trace):
 
     executed_task_info = scheduler.run(baseline, task_trace)
 
-    num_idle_cores_per_time_step = get_idle_core_count_from_task_info(
-        baseline,
-        executed_task_info,
-        task_trace,
-        num_vms,
-        num_cpus_per_vm,
-    )
-    for time_step in num_idle_cores_per_time_step:
-        write_line_to_csv(
+    # For granny we get the idle cores as we run the experiment, from the
+    # planner
+    if baseline in NATIVE_BASELINES:
+        num_idle_cores_per_time_step = get_idle_core_count_from_task_info(
             baseline,
-            IDLE_CORES_FILE_PREFIX,
+            executed_task_info,
+            task_trace,
             num_vms,
-            trace,
-            time_step,
-            num_idle_cores_per_time_step[time_step],
+            num_cpus_per_vm,
         )
+        for time_step in num_idle_cores_per_time_step:
+            write_line_to_csv(
+                baseline,
+                IDLE_CORES_FILE_PREFIX,
+                num_vms,
+                trace,
+                time_step,
+                num_idle_cores_per_time_step[time_step],
+            )
 
     # Finally shutdown the scheduler
     scheduler.shutdown()