Merge pull request #49 from sidkumar99/vfs_cache_pressure_tuning

adaptivemm: initial vm.vfs_cache_pressure tuning algorithm

Implement a simple vm.vfs_cache_pressure tuning algorithm. The algorithm works by using a user supplied hint which is the PREFER_OBJECT_CACHING option in the configuration file. This hint indicates if the user would rather have their system preserve dentry and inode object caches at the cost of higher memory utilizaiton. Some workloads are strongly impacted by the level on dentry and inode caching so it may benefit these workloads to have the mm system deprioritize reclaiming these caches. If this option is selected, the inode and dentry numbers are monitored and if they are projected to increase (via the slope of a best line fit), the sysctl is lowered by 10% and if they are projected to decrease, the sysctl is increased by 10%. The opposite behavior occurs if PREFER_OBJECT_CACHING is specified to 0.

Author: sidkumar99

adaptivemm/adaptivemmd.cfg

@@ -14,6 +14,9 @@ AGGRESSIVENESS=2
 # Enable management of free pages through watermarks tuning
 ENABLE_FREE_PAGE_MGMT=1
 
+# Guidance for tuning of vm.vfs_cache_pressure
+PREFER_OBJECT_CACHING=1
+
 # Maximum gap between low and high watermarks (in GB)
 # MAXGAP=5
 

adaptivemm/src/adaptivemmd.c

@@ -58,12 +58,15 @@
 #define KPAGEFLAGS		"/proc/kpageflags"
 #define MODULES			"/proc/modules"
 #define HUGEPAGESINFO		"/sys/kernel/mm/hugepages"
+#define DENTRYINFO		"/proc/sys/fs/dentry-state"
+#define INODEINFO		"/proc/sys/fs/inode-nr"
 
 /*
  * System files to control reclamation and compaction
  */
 #define RESCALE_WMARK		"/proc/sys/vm/watermark_scale_factor"
 #define	COMPACT_PATH_FORMAT	"/sys/devices/system/node/node%d/compact"
+#define VFS_CACHE_PRESSURE	"/proc/sys/vm/vfs_cache_pressure"
 
 /*
  * System files to control negative dentries
@@ -76,12 +79,16 @@
 #define CONFIG_FILE1		"/etc/sysconfig/adaptivemmd"
 #define CONFIG_FILE2		"/etc/default/adaptivemmd"
 
+#define FS_FIELDS	2
 #define MAX_NUMANODES	1024
 
 #define MAX_VERBOSE	5
 #define MAX_AGGRESSIVE	3
 #define MAX_NEGDENTRY	100
 
+#define MAX_VFS_CACHE_PRESSURE	1000
+#define MIN_VFS_CACHE_PRESSURE	3
+
 /*
  * Number of consecutive samples showing growth in unaccounted memory
  * that will trigger memory leak warning
@@ -91,17 +98,22 @@
 /* Minimum % change in meminfo numbers to trigger warnings */
 #define MEM_TRIGGER_DELTA	10
 
+int MAX(unsigned long a, unsigned long b) { return((a) > (b) ? a : b); }
+int MIN(unsigned long a, unsigned long b) { return((a) < (b) ? a : b); }
+
 unsigned long min_wmark[MAX_NUMANODES], low_wmark[MAX_NUMANODES];
 unsigned long high_wmark[MAX_NUMANODES], managed_pages[MAX_NUMANODES];
 unsigned long total_free_pages, total_cache_pages, total_hugepages, base_psize;
 long compaction_rate, reclaim_rate;
 struct lsq_struct page_lsq[MAX_NUMANODES][MAX_ORDER];
+struct lsq_struct fs_lsq[FS_FIELDS];
 int dry_run;
 int debug_mode, verbose, del_lock = 0;
 unsigned long maxgap;
 int aggressiveness = 2;
 int periodicity, skip_dmazone;
 int neg_dentry_pct = 15;	/* default is 1.5% */
+int prefer_object_caching = 1;
 
 /* Flags to enable various modules */
 bool memory_pressure_check_enabled = true;
@@ -127,7 +139,6 @@ unsigned int mywsf;
  */
 int max_compaction_order = MAX_ORDER - 4;
 
-
 /*
  * Clean up before exiting
  */
@@ -185,6 +196,14 @@ void log_msg(int level, char *fmt, ...)
 	va_end(args);
 }
 
+static inline unsigned long get_msecs(struct timespec *spec)
+{
+	if (!spec)
+		return -1;
+
+	return (unsigned long)((spec->tv_sec * 1000) + (spec->tv_nsec / 1000));
+}
+
 /*
  * Initiate memory compactiomn in the kernel on a given node.
  */
@@ -519,6 +538,95 @@ void rescale_maxwsf()
 		log_warn("Failed to compute reasonable WSF, %ld, total pages %ld, reclaimable pages %ld", new_wsf, total_managed, reclaimable_pages);
 }
 
+int parse_fs_files(char * path, unsigned long *data)
+{
+	FILE *fp = NULL;
+	char line[80];
+
+	fp = fopen(path, "r");
+	if (!fp)
+		return -1;
+
+	if (fgets(line, sizeof(line), fp) != NULL)
+		sscanf(line, "%lu\n", data);
+	else {
+		fclose(fp);
+		return -1;
+	}
+
+	fclose(fp);
+	return 0;
+}
+
+int set_vfs_cache_pressure(unsigned long new_cache_pressure)
+{
+	int fd;
+	char scaled_cache_pressure[20];
+
+	sprintf(scaled_cache_pressure, "%ld\n", new_cache_pressure);
+	if ((fd = open(VFS_CACHE_PRESSURE, O_WRONLY)) == -1) {
+		log_err("Failed to open "VFS_CACHE_PRESSURE" (%s)", strerror(errno));
+		return -1;
+	}
+
+	if (write(fd, scaled_cache_pressure, strlen(scaled_cache_pressure)) < 0) {
+		log_err("Failed to write to "VFS_CACHE_PRESSURE" (%s)", strerror(errno));
+		return -1;
+	}
+
+	log_info(1, "New vfs_cache_pressure = %ld", new_cache_pressure);
+	return 0;
+}
+
+/*
+ * rescale vm.vfs_cache_pressure value based on current memory trends
+ */
+void rescale_vfs_cache_pressure()
+{
+	struct timespec spec;
+	unsigned long curr_inodes;
+	unsigned long curr_dentries;
+	unsigned long curr_vfs_cache_pressure;
+	long long dentry_m, inode_m;
+	long long dentry_c, inode_c;
+
+	if (parse_fs_files(VFS_CACHE_PRESSURE, &curr_vfs_cache_pressure) ||
+		parse_fs_files(INODEINFO, &curr_inodes) ||
+		parse_fs_files(DENTRYINFO, &curr_dentries))
+			return;
+
+	clock_gettime(CLOCK_MONOTONIC_RAW, &spec);
+
+	if(lsq_fit(&fs_lsq[0], curr_dentries,
+				(long long)get_msecs(&spec),&dentry_m, &dentry_c) ||
+		lsq_fit(&fs_lsq[1], curr_inodes,
+				(long long)get_msecs(&spec), &inode_m, &inode_c))
+			return;
+	/*
+	 * adjust vm.vfs_cache_pressure based on the
+	 * growth of cached inode and dentry objects
+	 */
+	if (prefer_object_caching) {
+		if ((dentry_m > 0 || inode_m > 0) &&
+			curr_vfs_cache_pressure > MIN_VFS_CACHE_PRESSURE) {
+			set_vfs_cache_pressure(curr_vfs_cache_pressure * 0.9);
+		} else if ((dentry_m < 0 || inode_m < 0) &&
+			curr_vfs_cache_pressure < MAX_VFS_CACHE_PRESSURE)
+			set_vfs_cache_pressure(MIN(MAX_VFS_CACHE_PRESSURE,
+						curr_vfs_cache_pressure * 1.1));
+	} else {
+		if ((dentry_m > 0 || inode_m > 0) &&
+			curr_vfs_cache_pressure < MAX_VFS_CACHE_PRESSURE) {
+			set_vfs_cache_pressure(MIN(MAX_VFS_CACHE_PRESSURE,
+					curr_vfs_cache_pressure * 1.1));
+		} else if ((dentry_m < 0 || inode_m < 0) &&
+			curr_vfs_cache_pressure > MIN_VFS_CACHE_PRESSURE)
+			set_vfs_cache_pressure(curr_vfs_cache_pressure * 0.9);
+	}
+
+	return;
+}
+
 /*
  * Get the number of pages stolen by kswapd from /proc/vmstat.
  */
@@ -783,14 +891,6 @@ void rescale_watermarks(int scale_up)
 	close(fd);
 }
 
-static inline unsigned long get_msecs(struct timespec *spec)
-{
-	if (!spec)
-		return -1;
-
-	return (unsigned long)((spec->tv_sec * 1000) + (spec->tv_nsec / 1000));
-}
-
 /*
  * check_permissions() - Check all required permissions for this program to
  *			run successfully
@@ -1596,8 +1696,7 @@ void check_memory_pressure(bool init)
 
 /*
  * one_time_initializations() - Initialize settings that are set once at
- *	adaptivemmd startup
- *
+ * adaptivemmd startup
  */
 void one_time_initializations()
 {
@@ -1631,6 +1730,7 @@ void one_time_initializations()
 #define OPT_NEG_DENTRY1	"NEG-DENTRY-CAP"
 #define OPT_NEG_DENTRY2	"NEG_DENTRY_CAP"
 #define OPT_ENB_MEMLEAK	"ENABLE_MEMLEAK_CHECK"
+#define OPT_PREFER_OBJECT_CACHING "PREFER_OBJECT_CACHING"
 
 int parse_config()
 {
@@ -1728,6 +1828,8 @@ int parse_config()
 				neg_dentry_pct = val;
 		} else if (strncmp(token, OPT_ENB_MEMLEAK, sizeof(OPT_ENB_MEMLEAK)) == 0)
 			memleak_check_enabled = ((val==0)?false:true);
+		else if (strncmp(token, OPT_PREFER_OBJECT_CACHING, sizeof(OPT_PREFER_OBJECT_CACHING)) == 0)
+			prefer_object_caching = val;
 		else {
 			log_err("Error in configuration file at token \"%s\". Proceeding with defaults", token);
 			break;
@@ -1908,7 +2010,6 @@ int main(int argc, char **argv)
 	base_psize = getpagesize()/1024;
 
 	pr_info("adaptivemmd "VERSION" started (verbose=%d, aggressiveness=%d, maxgap=%d)", verbose, aggressiveness, maxgap);
-
 	one_time_initializations();
 
 	while (1) {
@@ -1927,9 +2028,9 @@ int main(int argc, char **argv)
 			updates_for_hugepages(retval);
 		if (maxgap == 0)
 			rescale_maxwsf();
-
 		check_memory_pressure(false);
 		check_memory_leak(false);
+		rescale_vfs_cache_pressure();
 
 		sleep(periodicity);
 	}

adaptivemm/src/predict.c

@@ -33,12 +33,13 @@
  * best.  The formulation is for the special case in which x_i = i + 1 - N;
  * this reduces the need for storage and permits constant time updates.
  */
-static int lsq_fit(struct lsq_struct *lsq, long long new_y, long long new_x,
+int lsq_fit(struct lsq_struct *lsq, long long new_y, long long new_x,
 	long long *m, long long *c)
 {
 	long long sigma_x, sigma_y;
 	long sigma_xy, sigma_xx;
 	long long slope_divisor;
+	long long numerator;
 	int i, next;
 	long x_offset;
 
@@ -97,7 +98,8 @@ static int lsq_fit(struct lsq_struct *lsq, long long new_y, long long new_x,
 	if (slope_divisor == 0)
 		return -1;
 
-	*m = ((LSQ_LOOKBACK * sigma_xy - sigma_x * sigma_y) * 100) / slope_divisor;
+	numerator = (LSQ_LOOKBACK * sigma_xy - sigma_x * sigma_y) * 100;
+	*m = numerator / slope_divisor;
 	*c = (sigma_y - *m * sigma_x) / LSQ_LOOKBACK;
 
 	/*
@@ -300,7 +302,7 @@ unsigned long predict(struct frag_info *frag_vec, struct lsq_struct *lsq,
 		 * graph.
 		 */
 		clock_gettime(CLOCK_MONOTONIC_RAW, &tspec);
-		current_time = tspec.tv_sec*1000 + tspec.tv_nsec/1000 - lsq->x[lsq->next];;
+		current_time = tspec.tv_sec*1000 + tspec.tv_nsec/1000 - lsq->x[lsq->next];
 		if (current_time < 0)
 			current_time = 0;
 		if ((x_cross < 0) ||

adaptivemm/src/predict.h

@@ -68,6 +68,10 @@ struct frag_info {
 	long long msecs;
 };
 
+
+int lsq_fit(struct lsq_struct *lsq, long long new_y, long long new_x,
+	long long *m, long long *c);
+
 unsigned long predict(struct frag_info *, struct lsq_struct *,
 			unsigned long, unsigned long, int);