locksmith.c: Added better cos swapping support

clients/square_evictions.c: Added support for locksmith histogram
rdtscp.h: Added fast log2 for square_evictions locksmith histogram array index calculation
prep_system.c: git operations will no longer open less

clients/square_evictions.c

@@ -18,6 +18,11 @@
 #define DEFAULT_PERCENT_CONTRACTED 80
 #define DEFAULT_PERCENT_EXPANDED   220
 
+#define MAX_MEMREF_CYCLES_POW	17
+#define LOWER(x) (1 << x)
+#define UPPER(x) (1 << (x+1))
+uint64_t histogram[MAX_MEMREF_CYCLES_POW] = {0};
+
 #define PERF_LOG_FILE_HEADER_LINE  "realtime,cputime,instructions,bandwidth\n"
 static const struct perf_event_attr PERF_LOG_COUNTERS[] = {
 	{
@@ -203,6 +208,10 @@ static int square_evictions(int cache_line_size, int num_periods, int passes_per
 						max_time = time;
 						printf("Current max cycles for a memory access %lu\n", max_time);
 					}
+					// Ignore the first 2 cycles due to memory fetch behavior
+					if(cycle < 2) {
+						++histogram[logtwo(time)];
+					}
 				}
 			}
 			assert(!siz || seen_initial);
@@ -429,5 +438,10 @@ int main(int argc, char *argv[]) {
 		munmap(perfbuf, perfbuf->siz);
 	if(perflog)
 		fclose(perflog);
+	if(measure_all) {
+		for(int i = 0; i < MAX_MEMREF_CYCLES_POW; ++i) {
+			printf("Mem access cycles >= 2^%d: %lu\n", i, histogram[i]);
+		}
+	}
 	return ret;
 }

locksmith.c

@@ -6,15 +6,20 @@
 #include "bench_commands.h"
 #include "log.h"
 #include "prep_system.h"
-#define LARGE_NUMBER "3"
+
+// Should be at least 2, on 1 it might measure memory fetches, 2+ will use definitely use cache
+#define LARGE_NUMBER "2"
 
 #define COMMON_COS 	1
 #define OTHER_COS 	2
 
 #define READER_CORE 	0
 #define SWITCHER_CORE	1
 
-#define NUM_SWITCHES	(1 << 25)
+// 21 takes about 8 seconds
+// 23 takes about 40 seconds
+// 24 takes about 1 minute
+#define NUM_SWITCHES	(1 << 24)
 
 // TODO add flag to square_evictions for timing each memory access
 static test_prog_t progs[] = {
@@ -42,8 +47,13 @@ static const struct pqos_l3ca COS_MASKS[] = {
 static const size_t NUM_COS_MASKS = sizeof COS_MASKS / sizeof *COS_MASKS;
 
 int main(int argc, char** argv) {
-	(void) argc;
-	(void) argv;
+	int swappy = 0;
+	/* lol */
+	if(argc > 1) {
+		if(argv[1][0] == '-' && argv[1][1] == 's')
+			swappy = 1;
+	}
+	/* end lol */
 	if(prep_system(false, -1) < 0) {
 		log_msg(LOG_FATAL, "Failed to complete initial setup\n");
 		return 1;
@@ -65,16 +75,17 @@ int main(int argc, char** argv) {
 		sleep(2);
 	}
 	run_benchmarks(progs, 1);
-	for(int i = 0; i < NUM_SWITCHES; ++i) {
-		if(pqos_l3ca_assoc_set(SWITCHER_CORE, OTHER_COS) != PQOS_RETVAL_OK) {
-			log_msg(LOG_FATAL, "Failed to switch Cos on switcher process\n");
+	if(swappy) {
+		printf("Swapping CoSes\n");
+		for(int i = 0; i < NUM_SWITCHES; ++i) {
+			if(pqos_l3ca_assoc_set(SWITCHER_CORE, OTHER_COS) != PQOS_RETVAL_OK) {
+				log_msg(LOG_FATAL, "Failed to switch Cos on switcher process\n");
+			}
 		}
+		/* Keep swapping COS of our other core */
+		// TODO fix this
+		printf("Done swapping!\n");
 	}
-	/* Keep swapping COS of our other core */
-	// TODO fix this
-	printf("Done swapping!\n");
-	wait_benchmarks();
-	run_benchmarks(progs, 1);
 	wait_benchmarks();
 	cleanup_system(true);
 }

prep_system.c

@@ -45,7 +45,7 @@ static bool rearrange_processes(bool multicore, int procs_go_where,
 
 #define GIT_LOG_CMD 	"git log --online -1"
 #define GIT_STATUS_CMD  "git status -uno"
-#define GIT_DIFF_CMD	"git diff HEAD"
+#define GIT_DIFF_CMD	"git --no-pager diff HEAD"
 #define DELIM 		"==="
 // Purposely ignoring the usual style for pointers!
 // I don't know enough about PL to argue about it though!

rdtscp.h

@@ -16,4 +16,12 @@ static inline uint64_t rdtscp(void) {
 	return ((uint64_t)cycles_high << 32 | cycles_low);
 }
 
+static inline uint64_t logtwo(const uint64_t x) {
+	uint64_t y;
+	__asm volatile( "\tbsr %1, %0\n"
+			: "=r"(y)
+			: "r" (x)
+	    );
+	return y;
+}
 #endif