Add a phenomenal script to distill SMT pair data to Mij and Mi scores

1 files changed, 291 insertions, 0 deletions

diff --git a/smt_analysis/computeSMTslowdown.py b/smt_analysis/computeSMTslowdown.py
new file mode 100755
index 0000000..2cf58ac
--- /dev/null
+++ b/smt_analysis/computeSMTslowdown.py
@@ -0,0 +1,291 @@
+#!/usr/bin/python3
+from typing import List, Any
+import numpy as np
+from scipy import stats
+import sys
+import os
+import plotille.plotille as plt
+TIMING_ERROR = 1000 #ns
+LEVEL_C_ANALYSIS = False
+
+def print_usage(argv):
+    print("This program takes in the all-pairs and baseline SMT data and computes how much each program is slowed when SMT in enabled.", file=sys.stderr)
+    print("Level-A/B usage: {} <file -A> <file -B> <baseline file> --cij".format(argv[0]), file=sys.stderr)
+    print("Level-C usage: {} <continuous pairs> <baseline file>".format(argv[0]), file=sys.stderr)
+
+# Check that we got the right number of parameters
+if len(sys.argv) < 3:
+    print_usage(sys.argv)
+    exit()
+
+if len(sys.argv) > 3:
+    print("Analyzing results using Level-A/B methodology...")
+else:
+    print("Analyzing results using Level-C methodology...")
+    LEVEL_C_ANALYSIS = True
+
+# Check that all input files are valid
+for f in sys.argv[1:-1]:
+    if not os.path.exists(f) or os.path.getsize(f) == 0:
+        print("ERROR: File '{}' does not exist or is empty".format(f), file=sys.stderr);
+        print_usage(sys.argv)
+        exit()
+
+# Print Cij values rather than Mij
+TIMES_ONLY = len(sys.argv) > 4 and "--cij" in sys.argv[4]
+OK_PAIRS_ONLY = len(sys.argv) > 4 and "--cij-ok" in sys.argv[4]
+
+# This parses the result data from unthreaded timing experiments
+# @param f File name to load
+# @returns res Map of benchmark name to sample count
+# @returns samples Map of benchmark name to list of execution time samples
+# @returns max_res May of benchmark to maximum execution time among all samples for that benchmark
+def load_baseline(f):
+    # constants for columns of baseline data files
+    TOTAL_NS = 5
+    BENCH_NAME = 0
+    SAMPLES = 4
+
+    # Load baseline data. This logic is based off the summarize programs
+    res = {} # Map of benchmark to list of all execution time samples
+    samples = {} # Map of benchmark name to sample count
+    max_res = {} # Map of benchmark name to maximum execution time
+
+    with open(f) as fp:
+        for line in fp:
+            s = line.split()
+            if s[BENCH_NAME] not in res:
+                res[s[BENCH_NAME]] = list([int(s[TOTAL_NS])])
+                samples[s[BENCH_NAME]] = int(s[SAMPLES])
+                max_res[s[BENCH_NAME]] = int(s[TOTAL_NS])
+            else:
+                res[s[BENCH_NAME]].append(int(s[TOTAL_NS]))
+                max_res[s[BENCH_NAME]] = max(int(s[TOTAL_NS]), max_res[s[BENCH_NAME]])
+    return res, samples, max_res
+
+# This parses the result data from paired, threaded timing experiements
+# @param file1 The -A file name
+# @param file2 The -B file name
+# @returns time 2D array of benchmark IDs to list of total container execution times
+# @returns offset 2D array of benchmark IDs to list of differences between the start
+#                 of the first and the start of the second benchmark
+# @returns name_to_idx Map of benchmark names to benchmark IDs
+# @returns idx_to_name List which when indexed with benchmark ID will yield the benchmark name
+def load_paired(file1, file2, benchmarkCount):
+    # constants for columns of paired data files
+    FIRST_PROG = 0
+    SECOND_PROG = 1
+    FIRST_CORE = 2
+    SECOND_CORE = 3
+    TRIALS = 4
+    START_S = 5 # Start seconds
+    START_N = 6 # Start nanoseconds
+    END_S = 7   # End seconds
+    END_N = 8   # End nanoseconds
+    RUN_ID = 9
+    JOB_NUM = 10
+
+    with open(file1) as f1:
+        numJobs = int(f1.readline().split()[TRIALS])
+    assert numJobs > 0
+    assert benchmarkCount > 0
+
+    # Total times of each container
+    time=[[[0 for x in range(numJobs)]for y in range(benchmarkCount)]for z in range(benchmarkCount)]
+    # Difference in time between when the first and the second task start in the container
+    offset=[[[0 for x in range(numJobs)]for y in range(benchmarkCount)]for z in range(benchmarkCount)]
+
+    # Some aggregate counters that we update as we go along
+    avg_off = 0
+    avg_off_samp = 0
+
+    # Load paired data
+    bench1 = 0 # Index to what's the current first benchmark being examined
+    bench2 = 0 # Index to what's the current second benchmark being examined
+
+    name_to_idx = {}
+    idx_to_name = [0 for x in range(benchmarkCount)]
+
+    job_idx = 0
+    with open(file1) as f1, open(file2) as f2:
+        for line1, line2 in zip(f1, f2):
+            lineArr1 = line1.split()
+            lineArr2 = line2.split()
+            start1 = int(lineArr1[START_S]) * 10**9 + int(lineArr1[START_N])
+            start2 = int(lineArr2[START_S]) * 10**9 + int(lineArr2[START_N])
+            minStart = min(start1, start2)
+            end1 = int(lineArr1[END_S]) * 10**9 + int(lineArr1[END_N])
+            end2 = int(lineArr2[END_S]) * 10**9 + int(lineArr2[END_N])
+            maxEnd = max(end1, end2)
+            # Time actually co-scheduled is minEnd - maxStart, but Sims uses a different model
+#            time[bench1][bench2][int(lineArr1[JOB_NUM])] = maxEnd - minStart
+            time[bench1][bench2][job_idx] = maxEnd - minStart
+            if lineArr1[SECOND_PROG] == "h264_dec" and lineArr2[JOB_NUM] == 0:
+                print(maxEnd - minStart)
+            # Compute offset: if first job starts at t=0, when does second start?
+#            offset[bench1][bench2][int(lineArr1[JOB_NUM])] = abs(start2-start1)
+            offset[bench1][bench2][job_idx] = abs(start2-start1)
+            # Compute some running statistics
+            avg_off += abs(start2-start1)
+            avg_off_samp += 1
+            # Increment to the next benchmark, this is weird because of the zip()
+            # This is doubly weird because our results are an upper trianguler matrix
+            if job_idx == numJobs - 1: #int(lineArr1[JOB_NUM]) == numJobs - 1:
+                if bench2 < benchmarkCount-1:
+                    bench2 = bench2 + 1
+                    job_idx = 0
+                else:
+                    name_to_idx[lineArr1[FIRST_PROG]] = bench1
+                    idx_to_name[bench1] = lineArr1[FIRST_PROG]
+                    bench1 = bench1 + 1
+                    bench2 = bench1 # bench1 will never again appear as bench2
+                    job_idx = 0
+            else:
+                job_idx += 1
+    print("Average offset is: " + str(avg_off/avg_off_samp) + "ns")
+    return time, offset, name_to_idx, idx_to_name
+
+# Pull in the data
+if not LEVEL_C_ANALYSIS:
+    baseline_times, baseline_sample_cnt, baseline_max_times = load_baseline(sys.argv[3])
+    paired_times, paired_offsets, name_to_idx, idx_to_name = load_paired(sys.argv[1], sys.argv[2], len(list(baseline_times.keys())))
+    for key in baseline_times:
+        print(key,max(baseline_times[key]))
+else:
+    baseline_times, baseline_sample_cnt, baseline_max_times = load_baseline(sys.argv[2])
+    # Paired times use an abuse of the baseline file format
+    paired_times_raw, _, _ = load_baseline(sys.argv[1])
+    benchmarkCount = int(np.sqrt(len(list(paired_times_raw.keys()))))
+    numJobs = len(next(iter(paired_times_raw.values())))
+    paired_times=[[[0 for x in range(numJobs)]for y in range(benchmarkCount)]for z in range(benchmarkCount)]
+    idx_to_name=[]
+    name_to_idx={}
+    bench1 = -1
+    #Generate the indexing approach
+    for pair in sorted(paired_times_raw.keys()):
+        [bench1name, bench2name] = pair.split('+') # Benchmark name is pair concatenated together with a '+' delimiter
+        if bench1 == -1 or bench1name != idx_to_name[-1]:
+            idx_to_name.append(bench1name)
+            name_to_idx[bench1name] = len(idx_to_name) - 1
+            bench1 += 1
+    # Populate the array
+    for bench1 in range(len(idx_to_name)):
+        for bench2 in range(len(idx_to_name)):
+            paired_times[bench1][bench2] = paired_times_raw[idx_to_name[bench1]+"+"+idx_to_name[bench2]]
+
+# We work iff the baseline was run for the same set of benchmarks as the pairs were
+if sorted(baseline_times.keys()) != sorted(name_to_idx.keys()):
+    print("ERROR: The baseline and paired experiments were over a different set of benchmarks!", file=sys.stderr)
+    print("Baseline keys:", baseline_times.keys(), file=sys.stderr)
+    print("Paired keys:", name_to_idx.keys(), file=sys.stderr)
+    exit();
+
+# Only consider benchmarks that are at least an order of magnitude longer than the timing error
+reliableNames = []
+for i in range(0, len(name_to_idx)):
+    benchmark = idx_to_name[i]
+    if min(baseline_times[benchmark]) > TIMING_ERROR * 10:
+        reliableNames.append(benchmark)
+
+# Compute SMT slowdown for each benchmark
+# Output format: table, each row is one benchmark and each column is one benchmark
+#                each cell is base1 + base2*m = pair solved for m, aka (pair - base1) / base2
+# Print table header
+print("Bench          ", end=" ")
+for name in reliableNames:
+    if not TIMES_ONLY: print("{:<10.10}".format(name), end=" ")
+    if TIMES_ONLY: print("{:<12.12}".format(name), end=" ")
+print()
+# Print rows
+sample_f = max # Change this to np.mean to use mean values in Mij generation
+M_vals = []
+for b1 in reliableNames:
+    if not TIMES_ONLY: print("{:<14.14}:".format(b1), end=" ")
+    if TIMES_ONLY: print("{:<14.14}:".format(b1), end=" ")
+    for b2 in reliableNames:
+        if not LEVEL_C_ANALYSIS:
+            Ci = max(sample_f(baseline_times[b1]), sample_f(baseline_times[b2]))
+            Cj = min(sample_f(baseline_times[b1]), sample_f(baseline_times[b2]))
+            Cij = sample_f(paired_times[name_to_idx[b1]][name_to_idx[b2]])
+            if False:
+                M = np.std(paired_times[name_to_idx[b1]][name_to_idx[b2]]) / np.mean(paired_times[name_to_idx[b1]][name_to_idx[b2]])
+            else:
+                M = (Cij - Ci) / Cj
+            if Cij and Cj * 10 > Ci: # We don't pair tasks with more than a 10x difference in length
+                M_vals.append(M)
+                if not TIMES_ONLY: print("{:>10.3}".format(M), end=" ")
+            else:
+                if not TIMES_ONLY: print("{:>10}".format("N/A"), end=" ")
+
+            if TIMES_ONLY and (not OK_PAIRS_ONLY or Cj * 10 > Ci):
+                print("{:>12}".format(Cij), end=" ")
+            elif OK_PAIRS_ONLY and Cj * 10 <= Ci:
+                print("{:>12}".format("0"), end=" ")
+
+        else:
+            time_with_smt = sample_f(paired_times[name_to_idx[b1]][name_to_idx[b2]])
+            time_wout_smt = sample_f(baseline_times[b1])
+            M = time_with_smt / time_wout_smt
+            M_vals.append(M)
+            print("{:>10.3}".format(M), end=" ")
+    print("")
+# Print some statistics about the distribution
+print("Average: {:>5.3} with standard deviation {:>5.3} using `{}`".format(np.mean(M_vals), np.std(M_vals), sample_f.__name__))
+Ms = np.asarray(M_vals, dtype=np.float32)
+if not LEVEL_C_ANALYSIS:
+    print(np.sum(Ms <= 0), "of", len(M_vals), "M_i:j values are at most zero -", 100*np.sum(Ms <= 0)/len(M_vals), "percent")
+    print(np.sum(Ms > 1), "of", len(M_vals), "M_i:j values are greater than one -", 100*np.sum(Ms > 1)/len(M_vals), "percent")
+    M_vals_to_plot = Ms[np.logical_and(Ms > 0, Ms <= 1)]
+else:
+    print(np.sum(Ms <= 1), "of", len(M_vals), "M_i:j values are at most one -", 100*np.sum(Ms <= 1)/len(M_vals), "percent")
+    print(np.sum(Ms > 2), "of", len(M_vals), "M_i:j values are greater than two -", 100*np.sum(Ms > 2)/len(M_vals), "percent")
+    M_vals_to_plot = Ms
+
+print("Using Sim's analysis, average: {:>5.3} with standard deviation {:>5.3} using `{}`".format(np.mean(list(M_vals_to_plot)), np.std(list(M_vals_to_plot)), sample_f.__name__))
+print(plt.hist(M_vals_to_plot, bins=10))
+
+##### BELOW TEXT IS OLD OFFSET CODE (patched) #####
+## This still works, but is hacky and deprecated ##
+## PearsonR doesn't work though                  ##
+if not LEVEL_C_ANALYSIS:
+    benchmarkNames = idx_to_name
+    benchmarkCount = len(benchmarkNames)
+    numJobs = len(paired_times[0][0])
+
+    reliableNames=["ndes", "cjpeg_wrbmp", "adpcm_enc", "cjpeg_transupp", "epic", "gsm_dec", "h264_dec", "huff_enc", "rijndael_enc", "rijndael_dec", "gsm_enc", "ammunition", "mpeg2"]
+
+    #stats.pearsonr(time[b1][b2], oList),
+
+    with open("weakRelPairs_offset.csv", mode="w+") as f3:
+        print("Benchmark1", "Benchmark2", "minOffset", "maxOffset", "meanOffset", "meddOffset", "stdOffset", "minLength", "maxLength", sep=",", file=f3)
+        for b1 in range (0, benchmarkCount):
+            for b2 in range (0, benchmarkCount):
+                if benchmarkNames[b1] in reliableNames and benchmarkNames[b2] in reliableNames:
+                    #exclude last job due to inccurate timing
+                    oList = paired_offsets[b1][b2][:numJobs-1]
+                    jList = paired_times[b1][b2][:numJobs-1]
+#                   plt.scatter(oList, jList)
+#                   plt.title(benchmarkNames[b1] + ", " + benchmarkNames[b2])
+#                   plt.show()
+#                   print(benchmarkNames[b1], benchmarkNames[b2], min(oList), max(oList), np.mean(oList), np.median(oList), np.std(oList), stats.pearsonr(jList, oList), stats.spearmanr(jList, oList),  sep=",", file=f3)
+                    print(benchmarkNames[b1], benchmarkNames[b2], min(oList), max(oList), np.mean(oList), np.median(oList), np.std(oList), min(jList), max(jList),  sep=",", file=f3)
+"""
+#with open("reliableGraphs.csv", mode="x") as f3:
+        for b1 in range(0, benchmarkCount):
+            for b2 in range(0, benchmarkCount):
+                if benchmarkNames[b1] in reliableNames and benchmarkNames[b2] in reliableNames:
+                    oList = offset[b1][b2][:numJobs - 1]
+                    jList=time[b1][b2][:numJobs-1]
+                    # offset, time scatterplot
+                    plt.scatter(oList, jList)
+                    plt.title(benchmarkNames[b1] + " " + benchmarkNames[b2] + " Offsets v. Time")
+                    plt.show()
+                    #time histogram
+                    #plt.hist(jList, bins=10)
+                    #plt.title(benchmarkNames[b1] + benchmarkNames[b2] + "Completion Times")
+                    #plt.show()
+                    #offset histogram
+                    #plt.hist(oList, bins=10)
+                    #plt.title(benchmarkNames[b1] + benchmarkNames[b2] + "Offsets")
+                    #plt.show()
+"""