BFS: update dfatool port; add eval scripts

9 files changed, 102 insertions, 19 deletions

diff --git a/BFS/Makefile b/BFS/Makefile
index 2e1dcfe..a4ea69d 100644
--- a/BFS/Makefile
+++ b/BFS/Makefile
@@ -5,8 +5,6 @@ COMMON_INCLUDES := support
 HOST_SOURCES := $(wildcard host/*.c)
 DPU_SOURCES := $(wildcard dpu/*.c)
 
-.PHONY: all clean test
-
 COMMON_FLAGS := -Wall -Wextra -g -I${COMMON_INCLUDES}
 HOST_FLAGS := ${COMMON_FLAGS} -std=c11 -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS}
 DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS}
@@ -33,3 +31,5 @@ clean:
 
 test: all
 	bin/host_code
+
+.PHONY: all clean test
diff --git a/BFS/baselines/cpu/Makefile b/BFS/baselines/cpu/Makefile
index 1b1d57c..6f082b1 100644
--- a/BFS/baselines/cpu/Makefile
+++ b/BFS/baselines/cpu/Makefile
@@ -10,6 +10,8 @@ bfs_O0: app.c
 bfs_O2: app.c
 	gcc -O2 -o bfs_O2 -fopenmp app.c
 
+# each bfs invocation performs 100 iterations
+
 .PHONY: run
 run: bfs
 	./bfs -f ../../data/loc-gowalla_edges.txt
diff --git a/BFS/baselines/cpu/app.c b/BFS/baselines/cpu/app.c
index 65cd420..caf4cbc 100644
--- a/BFS/baselines/cpu/app.c
+++ b/BFS/baselines/cpu/app.c
@@ -150,18 +150,14 @@ int main(int argc, char** argv) {
         }
 
         if (isOK) {
-            printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                "| throughput_cpu_ref_MBps=%f throughput_cpu_omp_MBps=%f\n",
+            printf("[::] BFS CPU | n_threads=%d e_type=%s n_elements=%d "
+                "| throughput_seq_MBps=%f throughput_MBps=%f",
                 nr_threads, "uint32_t", csrGraph.numNodes,
                 csrGraph.numNodes * sizeof(uint32_t) / timer.time[1],
                 csrGraph.numNodes * sizeof(uint32_t) / timer.time[0]);
-            printf("[::] n_threads=%d e_type=%s n_elements=%d "
-                "| throughput_cpu_ref_MOpps=%f throughput_cpu_omp_MOpps=%f\n",
-                nr_threads, "uint32_t", csrGraph.numNodes,
+            printf(" throughput_seq_MOpps=%f throughput_MOpps=%f",
                 csrGraph.numNodes / timer.time[1],
                 csrGraph.numNodes / timer.time[0]);
-            printf("[::] n_threads=%d e_type=%s n_elements=%d |",
-                nr_threads, "uint32_t", csrGraph.numNodes);
             printAll(&timer, 1);
         }
 
diff --git a/BFS/baselines/cpu/run-opti.sh b/BFS/baselines/cpu/run-opti.sh
new file mode 100755
index 0000000..0a3a4a3
--- /dev/null
+++ b/BFS/baselines/cpu/run-opti.sh
@@ -0,0 +1,11 @@
+#!/bin/sh
+
+HOST="$(hostname)"
+
+echo $HOST
+
+make clean
+
+make run_O0 | sed 's/CPU/CPU O0/' | tee "${HOST}-O0.txt"
+
+make run_O2 | sed 's/CPU/CPU O2/' | tee "${HOST}-O2.txt"
diff --git a/BFS/baselines/cpu/run.sh b/BFS/baselines/cpu/run.sh
index 4174fc3..d0e0400 100755
--- a/BFS/baselines/cpu/run.sh
+++ b/BFS/baselines/cpu/run.sh
@@ -2,6 +2,12 @@
 
 set -e
 
+HOST="$(hostname)"
+
+echo $HOST
+
+(
+
 echo "prim-benchmarks BFS CPU (dfatool edition)"
 echo "Started at $(date)"
 echo "Revision $(git describe --always)"
@@ -9,10 +15,13 @@ echo "Revision $(git describe --always)"
 # default threads: 4
 
 # input size depends on file -> strong scaling only
+# each BFS invocation performs 100 iterations
+# roadNet-CA appears to be too small; frequently gives inf throughput
 
-make
-for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
-	for f in loc-gowalla_edges roadNet-CA; do
+make -B verbose=1
+for nr_threads in 88 64 44 1 2 4 6 8 12 16 20 24 32; do
+	for f in loc-gowalla_edges; do # roadNet-CA; do
 		OMP_NUM_THREADS=${nr_threads} timeout --foreground -k 1m 30m ./bfs -f ../../data/${f}.txt || true
 	done
 done
+) | tee "${HOST}-explore.txt"
diff --git a/BFS/host/app.c b/BFS/host/app.c
index e4ed9b5..44154e2 100644
--- a/BFS/host/app.c
+++ b/BFS/host/app.c
@@ -297,18 +297,14 @@ int main(int argc, char** argv) {
     }
 
     if (status) {
-        printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%d "
-            "| throughput_pim_MBps=%f throughput_MBps=%f\n",
+        printf("[::] BFS NMC | n_dpus=%d n_tasklets=%d e_type=%s n_elements=%d "
+            "| throughput_pim_MBps=%f throughput_MBps=%f",
             numDPUs, NR_TASKLETS, "uint32_t", numNodes,
             numNodes * sizeof(uint32_t) / (timer.time[2]),
             numNodes * sizeof(uint32_t) / (timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]));
-        printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%d "
-            "| throughput_pim_MOpps=%f throughput_MOpps=%f\n",
-            numDPUs, NR_TASKLETS, "uint32_t", numNodes,
+        printf(" throughput_pim_MOpps=%f throughput_MOpps=%f",
             numNodes / (timer.time[2]),
             numNodes / (timer.time[0] + timer.time[1] + timer.time[2] + timer.time[3] + timer.time[4]));
-        printf("[::] n_dpus=%d n_tasklets=%d e_type=%s n_elements=%d |",
-            numDPUs, NR_TASKLETS, "uint32_t", numNodes);
         printAll(&timer, 4);
     }
 
diff --git a/BFS/run-paper-strong-full.sh b/BFS/run-paper-strong-full.sh
new file mode 100755
index 0000000..8d64d63
--- /dev/null
+++ b/BFS/run-paper-strong-full.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+set -e
+
+(
+
+echo "prim-benchmarks BFS strong-full (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_dpus in 256 512 1024 2048; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} verbose=1; then
+			# repetition is not part of upstream setup
+			for i in `seq 1 50`; do
+				timeout --foreground -k 1m 3m bin/host_code -f data/loc-gowalla_edges.txt || true
+			done
+		fi
+	done
+done
+) | tee log-paper-strong-full.txt
diff --git a/BFS/run-paper-strong-rank.sh b/BFS/run-paper-strong-rank.sh
new file mode 100755
index 0000000..0b20323
--- /dev/null
+++ b/BFS/run-paper-strong-rank.sh
@@ -0,0 +1,23 @@
+#!/bin/bash
+
+set -e
+
+(
+
+echo "prim-benchmarks BFS strong-rank (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# 256 and 512 are not part of upstream
+for nr_dpus in 256 512 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} verbose=1; then
+			# repetition is not part of upstream setup
+			for i in `seq 1 50`; do
+				timeout --foreground -k 1m 3m bin/host_code -f data/loc-gowalla_edges.txt || true
+			done
+		fi
+	done
+done
+) | tee log-paper-strong-rank.txt
diff --git a/BFS/run-paper-weak.sh b/BFS/run-paper-weak.sh
new file mode 100755
index 0000000..121758a
--- /dev/null
+++ b/BFS/run-paper-weak.sh
@@ -0,0 +1,24 @@
+#!/bin/bash
+
+set -e
+
+(
+
+echo "prim-benchmarks BFS weak (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# 256 and 512 are not part of upstream
+for nr_dpus in 256 512 1 4 16 64; do
+	for nr_tasklets in 1 2 4 8 16; do
+		echo
+		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} verbose=1; then
+			# repetition is not part of upstream setup
+			for i in `seq 1 50`; do
+				# upstream code uses some kind of generated rMat graphs, but does not provide instructions for reproduction
+				timeout --foreground -k 1m 3m bin/host_code -f data/loc-gowalla_edges.txt || true
+			done
+		fi
+	done
+done |
+) tee log-paper-weak.txt