aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/sched.c
diff options
context:
space:
mode:
authorakpm@osdl.org <akpm@osdl.org>2006-01-12 04:05:30 -0500
committerLinus Torvalds <torvalds@g5.osdl.org>2006-01-12 12:08:50 -0500
commit198e2f181163233b379dc7ce8a6d7516b84042e7 (patch)
treecc4067ca1c81034ba8d214b7ff4c39f2f5be66ee /kernel/sched.c
parent4dc7a0bbeb6882ad665e588e82fabe5bb4645f2f (diff)
[PATCH] scheduler cache-hot-autodetect
) From: Ingo Molnar <mingo@elte.hu> This is the latest version of the scheduler cache-hot-auto-tune patch. The first problem was that detection time scaled with O(N^2), which is unacceptable on larger SMP and NUMA systems. To solve this: - I've added a 'domain distance' function, which is used to cache measurement results. Each distance is only measured once. This means that e.g. on NUMA distances of 0, 1 and 2 might be measured, on HT distances 0 and 1, and on SMP distance 0 is measured. The code walks the domain tree to determine the distance, so it automatically follows whatever hierarchy an architecture sets up. This cuts down on the boot time significantly and removes the O(N^2) limit. The only assumption is that migration costs can be expressed as a function of domain distance - this covers the overwhelming majority of existing systems, and is a good guess even for more assymetric systems. [ People hacking systems that have assymetries that break this assumption (e.g. different CPU speeds) should experiment a bit with the cpu_distance() function. Adding a ->migration_distance factor to the domain structure would be one possible solution - but lets first see the problem systems, if they exist at all. Lets not overdesign. ] Another problem was that only a single cache-size was used for measuring the cost of migration, and most architectures didnt set that variable up. Furthermore, a single cache-size does not fit NUMA hierarchies with L3 caches and does not fit HT setups, where different CPUs will often have different 'effective cache sizes'. To solve this problem: - Instead of relying on a single cache-size provided by the platform and sticking to it, the code now auto-detects the 'effective migration cost' between two measured CPUs, via iterating through a wide range of cachesizes. The code searches for the maximum migration cost, which occurs when the working set of the test-workload falls just below the 'effective cache size'. I.e. real-life optimized search is done for the maximum migration cost, between two real CPUs. This, amongst other things, has the positive effect hat if e.g. two CPUs share a L2/L3 cache, a different (and accurate) migration cost will be found than between two CPUs on the same system that dont share any caches. (The reliable measurement of migration costs is tricky - see the source for details.) Furthermore i've added various boot-time options to override/tune migration behavior. Firstly, there's a blanket override for autodetection: migration_cost=1000,2000,3000 will override the depth 0/1/2 values with 1msec/2msec/3msec values. Secondly, there's a global factor that can be used to increase (or decrease) the autodetected values: migration_factor=120 will increase the autodetected values by 20%. This option is useful to tune things in a workload-dependent way - e.g. if a workload is cache-insensitive then CPU utilization can be maximized by specifying migration_factor=0. I've tested the autodetection code quite extensively on x86, on 3 P3/Xeon/2MB, and the autodetected values look pretty good: Dual Celeron (128K L2 cache): --------------------- migration cost matrix (max_cache_size: 131072, cpu: 467 MHz): --------------------- [00] [01] [00]: - 1.7(1) [01]: 1.7(1) - --------------------- cacheflush times [2]: 0.0 (0) 1.7 (1784008) --------------------- Here the slow memory subsystem dominates system performance, and even though caches are small, the migration cost is 1.7 msecs. Dual HT P4 (512K L2 cache): --------------------- migration cost matrix (max_cache_size: 524288, cpu: 2379 MHz): --------------------- [00] [01] [02] [03] [00]: - 0.4(1) 0.0(0) 0.4(1) [01]: 0.4(1) - 0.4(1) 0.0(0) [02]: 0.0(0) 0.4(1) - 0.4(1) [03]: 0.4(1) 0.0(0) 0.4(1) - --------------------- cacheflush times [2]: 0.0 (33900) 0.4 (448514) --------------------- Here it can be seen that there is no migration cost between two HT siblings (CPU#0/2 and CPU#1/3 are separate physical CPUs). A fast memory system makes inter-physical-CPU migration pretty cheap: 0.4 msecs. 8-way P3/Xeon [2MB L2 cache]: --------------------- migration cost matrix (max_cache_size: 2097152, cpu: 700 MHz): --------------------- [00] [01] [02] [03] [04] [05] [06] [07] [00]: - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) [01]: 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) [02]: 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) [03]: 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) 19.2(1) [04]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) 19.2(1) [05]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) 19.2(1) [06]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - 19.2(1) [07]: 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) 19.2(1) - --------------------- cacheflush times [2]: 0.0 (0) 19.2 (19281756) --------------------- This one has huge caches and a relatively slow memory subsystem - so the migration cost is 19 msecs. Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Ashok Raj <ashok.raj@intel.com> Signed-off-by: Ken Chen <kenneth.w.chen@intel.com> Cc: <wilder@us.ibm.com> Signed-off-by: John Hawkes <hawkes@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Diffstat (limited to 'kernel/sched.c')
-rw-r--r--kernel/sched.c468
1 files changed, 468 insertions, 0 deletions
diff --git a/kernel/sched.c b/kernel/sched.c
index c0c60c926d5e..98461de1ab65 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -34,6 +34,7 @@
34#include <linux/notifier.h> 34#include <linux/notifier.h>
35#include <linux/profile.h> 35#include <linux/profile.h>
36#include <linux/suspend.h> 36#include <linux/suspend.h>
37#include <linux/vmalloc.h>
37#include <linux/blkdev.h> 38#include <linux/blkdev.h>
38#include <linux/delay.h> 39#include <linux/delay.h>
39#include <linux/smp.h> 40#include <linux/smp.h>
@@ -5082,7 +5083,470 @@ static void init_sched_build_groups(struct sched_group groups[], cpumask_t span,
5082 5083
5083#define SD_NODES_PER_DOMAIN 16 5084#define SD_NODES_PER_DOMAIN 16
5084 5085
5086/*
5087 * Self-tuning task migration cost measurement between source and target CPUs.
5088 *
5089 * This is done by measuring the cost of manipulating buffers of varying
5090 * sizes. For a given buffer-size here are the steps that are taken:
5091 *
5092 * 1) the source CPU reads+dirties a shared buffer
5093 * 2) the target CPU reads+dirties the same shared buffer
5094 *
5095 * We measure how long they take, in the following 4 scenarios:
5096 *
5097 * - source: CPU1, target: CPU2 | cost1
5098 * - source: CPU2, target: CPU1 | cost2
5099 * - source: CPU1, target: CPU1 | cost3
5100 * - source: CPU2, target: CPU2 | cost4
5101 *
5102 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5103 * the cost of migration.
5104 *
5105 * We then start off from a small buffer-size and iterate up to larger
5106 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5107 * doing a maximum search for the cost. (The maximum cost for a migration
5108 * normally occurs when the working set size is around the effective cache
5109 * size.)
5110 */
5111#define SEARCH_SCOPE 2
5112#define MIN_CACHE_SIZE (64*1024U)
5113#define DEFAULT_CACHE_SIZE (5*1024*1024U)
5114#define ITERATIONS 2
5115#define SIZE_THRESH 130
5116#define COST_THRESH 130
5117
5118/*
5119 * The migration cost is a function of 'domain distance'. Domain
5120 * distance is the number of steps a CPU has to iterate down its
5121 * domain tree to share a domain with the other CPU. The farther
5122 * two CPUs are from each other, the larger the distance gets.
5123 *
5124 * Note that we use the distance only to cache measurement results,
5125 * the distance value is not used numerically otherwise. When two
5126 * CPUs have the same distance it is assumed that the migration
5127 * cost is the same. (this is a simplification but quite practical)
5128 */
5129#define MAX_DOMAIN_DISTANCE 32
5130
5131static unsigned long long migration_cost[MAX_DOMAIN_DISTANCE] =
5132 { [ 0 ... MAX_DOMAIN_DISTANCE-1 ] = -1LL };
5133
5134/*
5135 * Allow override of migration cost - in units of microseconds.
5136 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5137 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5138 */
5139static int __init migration_cost_setup(char *str)
5140{
5141 int ints[MAX_DOMAIN_DISTANCE+1], i;
5142
5143 str = get_options(str, ARRAY_SIZE(ints), ints);
5144
5145 printk("#ints: %d\n", ints[0]);
5146 for (i = 1; i <= ints[0]; i++) {
5147 migration_cost[i-1] = (unsigned long long)ints[i]*1000;
5148 printk("migration_cost[%d]: %Ld\n", i-1, migration_cost[i-1]);
5149 }
5150 return 1;
5151}
5152
5153__setup ("migration_cost=", migration_cost_setup);
5154
5155/*
5156 * Global multiplier (divisor) for migration-cutoff values,
5157 * in percentiles. E.g. use a value of 150 to get 1.5 times
5158 * longer cache-hot cutoff times.
5159 *
5160 * (We scale it from 100 to 128 to long long handling easier.)
5161 */
5162
5163#define MIGRATION_FACTOR_SCALE 128
5164
5165static unsigned int migration_factor = MIGRATION_FACTOR_SCALE;
5166
5167static int __init setup_migration_factor(char *str)
5168{
5169 get_option(&str, &migration_factor);
5170 migration_factor = migration_factor * MIGRATION_FACTOR_SCALE / 100;
5171 return 1;
5172}
5173
5174__setup("migration_factor=", setup_migration_factor);
5175
5176/*
5177 * Estimated distance of two CPUs, measured via the number of domains
5178 * we have to pass for the two CPUs to be in the same span:
5179 */
5180static unsigned long domain_distance(int cpu1, int cpu2)
5181{
5182 unsigned long distance = 0;
5183 struct sched_domain *sd;
5184
5185 for_each_domain(cpu1, sd) {
5186 WARN_ON(!cpu_isset(cpu1, sd->span));
5187 if (cpu_isset(cpu2, sd->span))
5188 return distance;
5189 distance++;
5190 }
5191 if (distance >= MAX_DOMAIN_DISTANCE) {
5192 WARN_ON(1);
5193 distance = MAX_DOMAIN_DISTANCE-1;
5194 }
5195
5196 return distance;
5197}
5198
5199static unsigned int migration_debug;
5200
5201static int __init setup_migration_debug(char *str)
5202{
5203 get_option(&str, &migration_debug);
5204 return 1;
5205}
5206
5207__setup("migration_debug=", setup_migration_debug);
5208
5209/*
5210 * Maximum cache-size that the scheduler should try to measure.
5211 * Architectures with larger caches should tune this up during
5212 * bootup. Gets used in the domain-setup code (i.e. during SMP
5213 * bootup).
5214 */
5215unsigned int max_cache_size;
5216
5217static int __init setup_max_cache_size(char *str)
5218{
5219 get_option(&str, &max_cache_size);
5220 return 1;
5221}
5222
5223__setup("max_cache_size=", setup_max_cache_size);
5224
5225/*
5226 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5227 * is the operation that is timed, so we try to generate unpredictable
5228 * cachemisses that still end up filling the L2 cache:
5229 */
5230static void touch_cache(void *__cache, unsigned long __size)
5231{
5232 unsigned long size = __size/sizeof(long), chunk1 = size/3,
5233 chunk2 = 2*size/3;
5234 unsigned long *cache = __cache;
5235 int i;
5236
5237 for (i = 0; i < size/6; i += 8) {
5238 switch (i % 6) {
5239 case 0: cache[i]++;
5240 case 1: cache[size-1-i]++;
5241 case 2: cache[chunk1-i]++;
5242 case 3: cache[chunk1+i]++;
5243 case 4: cache[chunk2-i]++;
5244 case 5: cache[chunk2+i]++;
5245 }
5246 }
5247}
5248
5249/*
5250 * Measure the cache-cost of one task migration. Returns in units of nsec.
5251 */
5252static unsigned long long measure_one(void *cache, unsigned long size,
5253 int source, int target)
5254{
5255 cpumask_t mask, saved_mask;
5256 unsigned long long t0, t1, t2, t3, cost;
5257
5258 saved_mask = current->cpus_allowed;
5259
5260 /*
5261 * Flush source caches to RAM and invalidate them:
5262 */
5263 sched_cacheflush();
5264
5265 /*
5266 * Migrate to the source CPU:
5267 */
5268 mask = cpumask_of_cpu(source);
5269 set_cpus_allowed(current, mask);
5270 WARN_ON(smp_processor_id() != source);
5271
5272 /*
5273 * Dirty the working set:
5274 */
5275 t0 = sched_clock();
5276 touch_cache(cache, size);
5277 t1 = sched_clock();
5278
5279 /*
5280 * Migrate to the target CPU, dirty the L2 cache and access
5281 * the shared buffer. (which represents the working set
5282 * of a migrated task.)
5283 */
5284 mask = cpumask_of_cpu(target);
5285 set_cpus_allowed(current, mask);
5286 WARN_ON(smp_processor_id() != target);
5287
5288 t2 = sched_clock();
5289 touch_cache(cache, size);
5290 t3 = sched_clock();
5291
5292 cost = t1-t0 + t3-t2;
5293
5294 if (migration_debug >= 2)
5295 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5296 source, target, t1-t0, t1-t0, t3-t2, cost);
5297 /*
5298 * Flush target caches to RAM and invalidate them:
5299 */
5300 sched_cacheflush();
5301
5302 set_cpus_allowed(current, saved_mask);
5303
5304 return cost;
5305}
5306
5307/*
5308 * Measure a series of task migrations and return the average
5309 * result. Since this code runs early during bootup the system
5310 * is 'undisturbed' and the average latency makes sense.
5311 *
5312 * The algorithm in essence auto-detects the relevant cache-size,
5313 * so it will properly detect different cachesizes for different
5314 * cache-hierarchies, depending on how the CPUs are connected.
5315 *
5316 * Architectures can prime the upper limit of the search range via
5317 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5318 */
5319static unsigned long long
5320measure_cost(int cpu1, int cpu2, void *cache, unsigned int size)
5321{
5322 unsigned long long cost1, cost2;
5323 int i;
5324
5325 /*
5326 * Measure the migration cost of 'size' bytes, over an
5327 * average of 10 runs:
5328 *
5329 * (We perturb the cache size by a small (0..4k)
5330 * value to compensate size/alignment related artifacts.
5331 * We also subtract the cost of the operation done on
5332 * the same CPU.)
5333 */
5334 cost1 = 0;
5335
5336 /*
5337 * dry run, to make sure we start off cache-cold on cpu1,
5338 * and to get any vmalloc pagefaults in advance:
5339 */
5340 measure_one(cache, size, cpu1, cpu2);
5341 for (i = 0; i < ITERATIONS; i++)
5342 cost1 += measure_one(cache, size - i*1024, cpu1, cpu2);
5343
5344 measure_one(cache, size, cpu2, cpu1);
5345 for (i = 0; i < ITERATIONS; i++)
5346 cost1 += measure_one(cache, size - i*1024, cpu2, cpu1);
5347
5348 /*
5349 * (We measure the non-migrating [cached] cost on both
5350 * cpu1 and cpu2, to handle CPUs with different speeds)
5351 */
5352 cost2 = 0;
5353
5354 measure_one(cache, size, cpu1, cpu1);
5355 for (i = 0; i < ITERATIONS; i++)
5356 cost2 += measure_one(cache, size - i*1024, cpu1, cpu1);
5357
5358 measure_one(cache, size, cpu2, cpu2);
5359 for (i = 0; i < ITERATIONS; i++)
5360 cost2 += measure_one(cache, size - i*1024, cpu2, cpu2);
5361
5362 /*
5363 * Get the per-iteration migration cost:
5364 */
5365 do_div(cost1, 2*ITERATIONS);
5366 do_div(cost2, 2*ITERATIONS);
5367
5368 return cost1 - cost2;
5369}
5370
5371static unsigned long long measure_migration_cost(int cpu1, int cpu2)
5372{
5373 unsigned long long max_cost = 0, fluct = 0, avg_fluct = 0;
5374 unsigned int max_size, size, size_found = 0;
5375 long long cost = 0, prev_cost;
5376 void *cache;
5377
5378 /*
5379 * Search from max_cache_size*5 down to 64K - the real relevant
5380 * cachesize has to lie somewhere inbetween.
5381 */
5382 if (max_cache_size) {
5383 max_size = max(max_cache_size * SEARCH_SCOPE, MIN_CACHE_SIZE);
5384 size = max(max_cache_size / SEARCH_SCOPE, MIN_CACHE_SIZE);
5385 } else {
5386 /*
5387 * Since we have no estimation about the relevant
5388 * search range
5389 */
5390 max_size = DEFAULT_CACHE_SIZE * SEARCH_SCOPE;
5391 size = MIN_CACHE_SIZE;
5392 }
5393
5394 if (!cpu_online(cpu1) || !cpu_online(cpu2)) {
5395 printk("cpu %d and %d not both online!\n", cpu1, cpu2);
5396 return 0;
5397 }
5398
5399 /*
5400 * Allocate the working set:
5401 */
5402 cache = vmalloc(max_size);
5403 if (!cache) {
5404 printk("could not vmalloc %d bytes for cache!\n", 2*max_size);
5405 return 1000000; // return 1 msec on very small boxen
5406 }
5407
5408 while (size <= max_size) {
5409 prev_cost = cost;
5410 cost = measure_cost(cpu1, cpu2, cache, size);
5411
5412 /*
5413 * Update the max:
5414 */
5415 if (cost > 0) {
5416 if (max_cost < cost) {
5417 max_cost = cost;
5418 size_found = size;
5419 }
5420 }
5421 /*
5422 * Calculate average fluctuation, we use this to prevent
5423 * noise from triggering an early break out of the loop:
5424 */
5425 fluct = abs(cost - prev_cost);
5426 avg_fluct = (avg_fluct + fluct)/2;
5427
5428 if (migration_debug)
5429 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): (%8Ld %8Ld)\n",
5430 cpu1, cpu2, size,
5431 (long)cost / 1000000,
5432 ((long)cost / 100000) % 10,
5433 (long)max_cost / 1000000,
5434 ((long)max_cost / 100000) % 10,
5435 domain_distance(cpu1, cpu2),
5436 cost, avg_fluct);
5437
5438 /*
5439 * If we iterated at least 20% past the previous maximum,
5440 * and the cost has dropped by more than 20% already,
5441 * (taking fluctuations into account) then we assume to
5442 * have found the maximum and break out of the loop early:
5443 */
5444 if (size_found && (size*100 > size_found*SIZE_THRESH))
5445 if (cost+avg_fluct <= 0 ||
5446 max_cost*100 > (cost+avg_fluct)*COST_THRESH) {
5447
5448 if (migration_debug)
5449 printk("-> found max.\n");
5450 break;
5451 }
5452 /*
5453 * Increase the cachesize in 5% steps:
5454 */
5455 size = size * 20 / 19;
5456 }
5457
5458 if (migration_debug)
5459 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
5460 cpu1, cpu2, size_found, max_cost);
5461
5462 vfree(cache);
5463
5464 /*
5465 * A task is considered 'cache cold' if at least 2 times
5466 * the worst-case cost of migration has passed.
5467 *
5468 * (this limit is only listened to if the load-balancing
5469 * situation is 'nice' - if there is a large imbalance we
5470 * ignore it for the sake of CPU utilization and
5471 * processing fairness.)
5472 */
5473 return 2 * max_cost * migration_factor / MIGRATION_FACTOR_SCALE;
5474}
5475
5476static void calibrate_migration_costs(const cpumask_t *cpu_map)
5477{
5478 int cpu1 = -1, cpu2 = -1, cpu, orig_cpu = raw_smp_processor_id();
5479 unsigned long j0, j1, distance, max_distance = 0;
5480 struct sched_domain *sd;
5481
5482 j0 = jiffies;
5483
5484 /*
5485 * First pass - calculate the cacheflush times:
5486 */
5487 for_each_cpu_mask(cpu1, *cpu_map) {
5488 for_each_cpu_mask(cpu2, *cpu_map) {
5489 if (cpu1 == cpu2)
5490 continue;
5491 distance = domain_distance(cpu1, cpu2);
5492 max_distance = max(max_distance, distance);
5493 /*
5494 * No result cached yet?
5495 */
5496 if (migration_cost[distance] == -1LL)
5497 migration_cost[distance] =
5498 measure_migration_cost(cpu1, cpu2);
5499 }
5500 }
5501 /*
5502 * Second pass - update the sched domain hierarchy with
5503 * the new cache-hot-time estimations:
5504 */
5505 for_each_cpu_mask(cpu, *cpu_map) {
5506 distance = 0;
5507 for_each_domain(cpu, sd) {
5508 sd->cache_hot_time = migration_cost[distance];
5509 distance++;
5510 }
5511 }
5512 /*
5513 * Print the matrix:
5514 */
5515 if (migration_debug)
5516 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
5517 max_cache_size,
5518#ifdef CONFIG_X86
5519 cpu_khz/1000
5520#else
5521 -1
5522#endif
5523 );
5524 printk("migration_cost=");
5525 for (distance = 0; distance <= max_distance; distance++) {
5526 if (distance)
5527 printk(",");
5528 printk("%ld", (long)migration_cost[distance] / 1000);
5529 }
5530 printk("\n");
5531 j1 = jiffies;
5532 if (migration_debug)
5533 printk("migration: %ld seconds\n", (j1-j0)/HZ);
5534
5535 /*
5536 * Move back to the original CPU. NUMA-Q gets confused
5537 * if we migrate to another quad during bootup.
5538 */
5539 if (raw_smp_processor_id() != orig_cpu) {
5540 cpumask_t mask = cpumask_of_cpu(orig_cpu),
5541 saved_mask = current->cpus_allowed;
5542
5543 set_cpus_allowed(current, mask);
5544 set_cpus_allowed(current, saved_mask);
5545 }
5546}
5547
5085#ifdef CONFIG_NUMA 5548#ifdef CONFIG_NUMA
5549
5086/** 5550/**
5087 * find_next_best_node - find the next node to include in a sched_domain 5551 * find_next_best_node - find the next node to include in a sched_domain
5088 * @node: node whose sched_domain we're building 5552 * @node: node whose sched_domain we're building
@@ -5448,6 +5912,10 @@ next_sg:
5448#endif 5912#endif
5449 cpu_attach_domain(sd, i); 5913 cpu_attach_domain(sd, i);
5450 } 5914 }
5915 /*
5916 * Tune cache-hot values:
5917 */
5918 calibrate_migration_costs(cpu_map);
5451} 5919}
5452/* 5920/*
5453 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 5921 * Set up scheduler domains and groups. Callers must hold the hotplug lock.