Index: linux-2.6.23/include/linux/genetic.h =================================================================== --- /dev/null +++ linux-2.6.23/include/linux/genetic.h @@ -0,0 +1,285 @@ +#ifndef __LINUX_GENETIC_H +#define __LINUX_GENETIC_H +/* + * include/linux/genetic.h + * + * Jake Moilanen + * Copyright (C) 2004 IBM + * + * Genetic algorithm library + * + * This program is free software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as published + * by the Free Software Foundation. + */ + +#include +#include +#include +#include + + +#define GENETIC_HISTORY_SIZE 0x8 +#define GENETIC_HISTORY_MASK (GENETIC_HISTORY_SIZE - 1) + +/* percentage of total number genes to mutate */ +#define GENETIC_DEFAULT_MUTATION_RATE 15 + +/* XXX TODO Make this an adjustable runtime variable */ +/* Percentage that an iteration can jump within the range */ +#define GENETIC_ITERATIVE_MUTATION_RANGE 20 + +/* the rate that GENETIC_DEFAULT_MUTATION_RATE itself can change */ +#define GENETIC_DEFAULT_MUTATION_RATE_CHANGE 4 +#define GENETIC_MAX_MUTATION_RATE 45 +#define GENETIC_MIN_MUTATION_RATE 10 + +#define GENETIC_DEBUG 0 + +#ifdef CONFIG_FINGERPRINTING +#define FP_DECAY 90 +#define GENETIC_NUM_DEBUG_POINTS 5 +#else +#define GENETIC_NUM_DEBUG_POINTS 4 +#endif + +#define GENETIC_PRINT_DEBUG 0 +#define gen_dbg(format, arg...) do { if (GENETIC_PRINT_DEBUG) printk(KERN_EMERG __FILE__ ": " format "\n" , ## arg); } while (0) +#define gen_trc(format, arg...) do { if (GENETIC_PRINT_DEBUG) printk(KERN_EMERG __FILE__ ":%s:%d\n" , __FUNCTION__, __LINE__); } while (0) + +struct gene_param_s; +struct genetic_s; +struct phenotype_s; + +struct genetic_child_s { + struct list_head list; + long long fitness; + unsigned long num_genes; + void *genes; + struct gene_param_s *gene_param; + void *stats_snapshot; + int id; +}; + +typedef struct genetic_child_s genetic_child_t; + +/* Here's a generic idea of what it the genes could look like */ +struct gene_param_s { + char *name; + unsigned long min; + unsigned long max; + unsigned long initial; + void (*mutate_gene)(genetic_child_t *, unsigned long); +}; + +typedef struct gene_param_s gene_param_t; + +struct phenotype_s { + struct list_head phenotype; + + struct list_head children_queue[2]; + struct list_head *run_queue; + struct list_head *finished_queue; + struct genetic_ops *ops; + + char *name; + + struct genetic_s *genetic; /* point back + * to genetic + * struct + */ + + unsigned long num_children; /* Must be power of 2 */ + unsigned long natural_selection_cutoff; /* How many children + * will survive + */ + void *stats_snapshot; + unsigned long child_number; + + /* percentage of total number of genes to mutate */ + long mutation_rate; + unsigned long num_mutations; + unsigned long num_genes; + + genetic_child_t **child_ranking; + + void (*natural_selection)(struct phenotype_s *); + + /* This UID is bitmap comprised of other phenotypes that contribute + to the genes */ + unsigned long uid; + + /* performance metrics */ + long long avg_fitness; + long long last_gen_avg_fitness; + + unsigned long fitness_history_index; + long long fitness_history[GENETIC_HISTORY_SIZE]; + +#if GENETIC_DEBUG + unsigned long debug_size; /* number of longs in + debug history */ + unsigned long debug_index; + long long *debug_history; +#endif +#ifdef CONFIG_FINGERPRINTING + struct dentry *fp_dir; + struct fingerprint *fp; + struct fp_snapshot *fp_ss; + unsigned long ***top_child; + long long ***top_fitness; + int last_fingerprint; +#else + long long top_fitness; +#endif + + long long from_top; + +}; + +typedef struct phenotype_s phenotype_t; + +/** + * struct genetic_s - contains all data structures for a genetic plugin + * @name: string that will identify this genetic alg. in debugfs and printk + * @phenotype: list of all registered phenotypes + * @child_number: the running child index (< @num_children) + * @child_life_time: time in ms each child is ran before being swapped + * @num_children: number of children in each generation (must be a power of 2) + * @generation_number: increased once every child in a generation is ran + * @defaults: when 1 the genetic library will hold all genes at defaults + * @fingerprinting: when 1 the genetic library wil use gene fingerprinting if CONFIG_FINGERPRINTING + */ +struct genetic_s { + char *name; + struct timer_list timer; + + struct list_head phenotype; + + unsigned long child_number; + unsigned long child_life_time; + unsigned long num_children; + + unsigned long generation_number; + int defaults; + + /* private */ + struct dentry *dir; + struct dentry *phenotypes_dir; + struct dentry *fingerprinting_dir; +#ifdef CONFIG_FINGERPRINTING + int fingerprinting; +#endif +}; + +typedef struct genetic_s genetic_t; + +struct genetic_ops { + void (*create_child)(genetic_child_t *); + void (*set_child_genes)(void *); + void (*calc_fitness)(genetic_child_t *); + void (*combine_genes)(genetic_child_t *, genetic_child_t *, + genetic_child_t *); + void (*mutate_child)(genetic_child_t *); + void (*calc_post_fitness)(phenotype_t *); /* Fitness routine used when + * need to take into account + * other phenotype fitness + * results after they ran + */ + void (*take_snapshot)(phenotype_t *); + void (*shift_mutation_rate)(phenotype_t *); + int (*gene_show)(struct seq_file *, void *); +#ifdef CONFIG_FINGERPRINTING + void (*get_fingerprint)(phenotype_t *); + void (*update_fingerprint)(phenotype_t *); + void * (*create_top_genes)(phenotype_t *); + int (*top_fitness_show)(struct seq_file *, void *); + int (*snapshot_show)(struct seq_file *, void *); + int (*state_show)(struct seq_file *, void *); +#endif +}; + +/* Setup routines */ +int __init genetic_init(genetic_t ** in_genetic, unsigned long num_children, + unsigned long child_life_time, int fingerprinting, + char * name); +int __init genetic_register_phenotype(genetic_t * genetic, struct genetic_ops * ops, + unsigned long num_children, char * name, + unsigned long num_genes, unsigned long uid); +void __init genetic_start(genetic_t * genetic); + +/* Generic helper functions */ +void genetic_generic_mutate_child(genetic_child_t * child); +void genetic_generic_iterative_mutate_gene(genetic_child_t * child, unsigned long gene_num); +void genetic_generic_combine_genes(genetic_child_t * parent_a, + genetic_child_t * parent_b, + genetic_child_t * child); +void genetic_create_child_spread(genetic_child_t * child, unsigned long num_children); +void genetic_create_child_defaults(genetic_child_t * child); +void genetic_general_shift_mutation_rate(phenotype_t * in_pt); +int genetic_generic_gene_show(struct seq_file *s, void *unused); + + +/* XXX do this more intelligently */ +#ifndef DIVLL_OP +#define DIVLL_OP +#if BITS_PER_LONG >= 64 + +static inline void divll(long long *n, long div, long *rem) +{ + *rem = *n % div; + *n /= div; +} + +#else + +static inline void divl(int32_t high, int32_t low, + int32_t div, + int32_t *q, int32_t *r) +{ + int64_t n = (u_int64_t)high << 32 | low; + int64_t d = (u_int64_t)div << 31; + int32_t q1 = 0; + int c = 32; + while (n > 0xffffffff) { + q1 <<= 1; + if (n >= d) { + n -= d; + q1 |= 1; + } + d >>= 1; + c--; + } + q1 <<= c; + if (n) { + low = n; + *q = q1 | (low / div); + *r = low % div; + } else { + *r = 0; + *q = q1; + } + return; +} + +static inline void divll(long long *n, long div, long *rem) +{ + int32_t low, high; + low = *n & 0xffffffff; + high = *n >> 32; + if (high) { + int32_t high1 = high % div; + int32_t low1 = low; + high /= div; + divl(high1, low1, div, &low, (int32_t *)rem); + *n = (int64_t)high << 32 | low; + } else { + *n = low / div; + *rem = low % div; + } +} +#endif + +#endif /* #ifndef divll */ + +#endif Index: linux-2.6.23/lib/Kconfig =================================================================== --- linux-2.6.23.orig/lib/Kconfig +++ linux-2.6.23/lib/Kconfig @@ -63,6 +63,12 @@ config AUDIT_GENERIC depends on AUDIT && !AUDIT_ARCH default y +config GENETIC_LIB + bool "Genetic Library" + help + This option will build in a genetic library that will tweak + kernel parameters autonomically to improve performance. + # # compression support is select'ed if needed # Index: linux-2.6.23/lib/Makefile =================================================================== --- linux-2.6.23.orig/lib/Makefile +++ linux-2.6.23/lib/Makefile @@ -47,6 +47,7 @@ obj-$(CONFIG_CRC32) += crc32.o obj-$(CONFIG_CRC7) += crc7.o obj-$(CONFIG_LIBCRC32C) += libcrc32c.o obj-$(CONFIG_GENERIC_ALLOCATOR) += genalloc.o +obj-$(CONFIG_GENETIC_LIB) += genetic.o obj-$(CONFIG_ZLIB_INFLATE) += zlib_inflate/ obj-$(CONFIG_ZLIB_DEFLATE) += zlib_deflate/ Index: linux-2.6.23/lib/genetic.c =================================================================== --- /dev/null +++ linux-2.6.23/lib/genetic.c @@ -0,0 +1,892 @@ +/* + * Genetic Algorithm Library + * + * Jake Moilanen + * Copyright (C) 2004-2005 IBM + * + * This program is free software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as published + * by the Free Software Foundation. + */ + +/* + * Life cycle + * + * 1.) Create random children + * 2.) Run tests + * 3.) Calculate fitness + * 4.) Take top preformers + * 5.) Make children + * 6.) Mutate + * 7.) Goto step 2 + */ + +/* + * TODO: + * + * - Check to make sure DEF_DESKTOP_TIMESLICE is operating correctly + * - fix fixup_timeslice + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "genetic-debug.c" + +#ifdef CONFIG_FINGERPRINTING +#include +#include "fingerprinting.c" +#endif + +char genetic_lib_version[] = "0.3.1"; + +int mutation_rate_change = GENETIC_DEFAULT_MUTATION_RATE_CHANGE; +int genetic_lib_enabled = 1; + +static void genetic_ns_top_parents(phenotype_t *); +static void genetic_ns_award_top_parents(phenotype_t *); +static int genetic_create_children(phenotype_t *); +static void genetic_split_performers(phenotype_t *); +static void genetic_mutate(phenotype_t *); +static void genetic_run_child(genetic_t * genetic); +static void genetic_new_generation(genetic_t * genetic); + +void genetic_switch_child(unsigned long data); + + +int __init genetic_init(genetic_t ** in_genetic, unsigned long num_children, + unsigned long child_life_time, int fingerprinting, + char * name) +{ + genetic_t * genetic; + + if (!genetic_lib_enabled) + return 0; + + printk(KERN_INFO "Initializing Genetic Library - version %s\n", + genetic_lib_version); + + genetic = (genetic_t *)kmalloc(sizeof(genetic_t), GFP_KERNEL); + if (!genetic) { + printk(KERN_ERR "genetic_init: not enough memory\n"); + return -ENOMEM; + } + + *in_genetic = genetic; + + genetic->name = (char *)kmalloc(strlen(name), GFP_KERNEL); + if (!genetic->name) { + printk(KERN_ERR "genetic_init: not enough memory\n"); + kfree(genetic); + return -ENOMEM; + } + + /* Init some of our values */ + strcpy(genetic->name, name); + + genetic->num_children = num_children; + genetic->child_life_time = child_life_time; + + genetic->generation_number = 1; + genetic->child_number = 0; + genetic->defaults = 0; +#ifdef CONFIG_FINGERPRINTING + genetic->fingerprinting = fingerprinting; +#endif + + /* Setup how long each child has to live */ + init_timer(&genetic->timer); + genetic->timer.function = genetic_switch_child; + genetic->timer.data = (unsigned long)genetic; + + INIT_LIST_HEAD(&genetic->phenotype); + + /* Setup debugfs */ + genetic->dir = genetic_create_tree(name, NULL); + genetic->phenotypes_dir = genetic_create_tree("phenotypes", genetic->dir); + +#ifdef CONFIG_FINGERPRINTING + if (fingerprinting) + genetic->fingerprinting_dir = genetic_create_tree("fingerprinting", genetic->dir); +#endif + + /* TODO add stack to the genetic track dentries for deallocation */ + debugfs_create_file("stats", S_IFREG|S_IRUGO, genetic->dir, + genetic, &genetic_stat_operations); + + debugfs_create_file("phenotype_average", S_IFREG|S_IRUGO, genetic->dir, + genetic, &genetic_phenotype_average_operations); + + debugfs_create_bool("defaults", S_IWUSR|S_IFREG|S_IRUGO, genetic->dir, + &genetic->defaults); + + return 0; +} + +int __init genetic_register_phenotype(genetic_t * genetic, + struct genetic_ops * ops, unsigned long num_children, + char * name, unsigned long num_genes, unsigned long uid) +{ + phenotype_t * pt; + int rc; + + if (!genetic_lib_enabled) + return 0; + + printk(KERN_INFO "Initializing %s's phenotype %s\n", genetic->name, + name); + + pt = (phenotype_t *)kmalloc(sizeof(phenotype_t), GFP_KERNEL); + if (!genetic) { + printk(KERN_ERR "genetic_register_phenotype: not enough\ + memory\n"); + return -ENOMEM; + } + + pt->name = (char *)kmalloc(strlen(name), GFP_KERNEL); + if (!pt->name) { + printk(KERN_ERR "genetic_register_phenotype: not enough\ + memory\n"); + kfree(pt); + return -ENOMEM; + } + + pt->child_ranking = (genetic_child_t **)kmalloc(num_children * sizeof(genetic_child_t *), GFP_KERNEL); + if (!pt->child_ranking) { + printk(KERN_ERR "genetic_register_phenotype: not enough\ + memory\n"); + kfree(pt->name); + kfree(pt); + return -ENOMEM; + } + + + strcpy(pt->name, name); + + INIT_LIST_HEAD(&pt->children_queue[0]); + INIT_LIST_HEAD(&pt->children_queue[1]); + + pt->run_queue = &pt->children_queue[0]; + pt->finished_queue = &pt->children_queue[1]; + + pt->ops = ops; + pt->num_children = num_children; + + pt->mutation_rate = GENETIC_DEFAULT_MUTATION_RATE; + pt->natural_selection = genetic_ns_top_parents; + pt->natural_selection_cutoff = num_children / 2; + pt->avg_fitness = 0; + pt->last_gen_avg_fitness = 0; + pt->child_number = 0; + + pt->genetic = genetic; + pt->uid = uid; + pt->num_genes = num_genes; + + pt->top_fitness = 0; + +#ifdef CONFIG_FINGERPRINTING + if (genetic->fingerprinting) { + if ((rc = genetic_init_fingerprinting(pt)) < 0) + return rc; + } +#endif + + /* Create some children */ + rc = genetic_create_children(pt); + if (rc) + return rc; + + list_add_tail(&pt->phenotype, &genetic->phenotype); + + if (ops->gene_show) { + debugfs_create_file(name, S_IFREG|S_IRUGO, + genetic->phenotypes_dir, pt, &genetic_gene_operations); + } + + return 0; +} + + + +void __init genetic_start(genetic_t * genetic) +{ + if (!genetic_lib_enabled) + return; + + genetic_run_child(genetic); + printk(KERN_INFO "%ld children started in %s genetic library\n", + genetic->num_children, genetic->name); +} + + + +/* create some children, it is up to the lib user to come up w/ a good + distro of genes for it's children */ +static int genetic_create_children(phenotype_t * pt) +{ + unsigned long i; + genetic_child_t * child; + + for (i = 0; i < pt->num_children; i++) { + pt->child_ranking[i] = (genetic_child_t *)kmalloc( + sizeof(genetic_child_t), GFP_KERNEL); + + if (!pt->child_ranking[i]) { + printk(KERN_ERR "genetic_create_child: not enough\ + memory\n"); + + for (i = i - 1; i >= 0; i--) + kfree(pt->child_ranking[i]); + + return -ENOMEM; + } + + child = pt->child_ranking[i]; + + child->id = i; + + pt->ops->create_child(child); + + list_add_tail(&child->list, pt->run_queue); + } + + return 0; +} + + +/* See how well child did and run the next one */ +void genetic_switch_child(unsigned long data) +{ + genetic_t * genetic = (genetic_t *)data; + genetic_child_t * child; + + struct list_head * p; + phenotype_t * pt; + + int new_generation = 0; +#ifdef GENETIC_DEBUG_VERBOSE + printk(KERN_INFO "genetic_switch_child() for %s\n", genetic->name); +#endif + list_for_each(p, &genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + + child = list_entry(pt->run_queue->next, genetic_child_t, list); + +#ifdef GENETIC_DEBUG_VERBOSE + printk(KERN_INFO " phenotype %s\n", pt->name); +#endif + + list_del(&child->list); + + list_add_tail(&child->list, pt->finished_queue); + + if (pt->ops->calc_fitness) + pt->ops->calc_fitness(child); + +#ifdef GENETIC_DEBUG_VERBOSE + printk(KERN_INFO " finished calc_fitness\n"); +#endif + + pt->child_ranking[pt->child_number++] = child; + + /* See if need more children */ + if (list_empty(pt->run_queue)) + new_generation = 1; + + } + + genetic->child_number++; + + if (new_generation) + genetic_new_generation(genetic); + + genetic_run_child(genetic); + +#ifdef GENETIC_DEBUG_VERBOSE + printk("exiting genetic_switch_child()\n"); +#endif +} + +/* Set the childs genes for run */ +void genetic_run_child(genetic_t * genetic) +{ + struct list_head * p; + phenotype_t * pt; + + genetic_child_t * child; + void * genes; + + list_for_each(p, &genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + + child = list_entry(pt->run_queue->next, genetic_child_t, list); + + /* genetic alg. disabled, only use default genes */ + if (genetic->defaults) + genetic_create_child_defaults(child); + + genes = child->genes; + + if (pt->ops->set_child_genes) + pt->ops->set_child_genes(genes); + + if (pt->ops->take_snapshot) + pt->ops->take_snapshot(pt); + + } + + /* set a timer interrupt */ + genetic->timer.expires = jiffies + genetic->child_life_time; + add_timer(&genetic->timer); + +} + +/* This natural selection routine will take the top + * natural_select_cutoff and use them to make children for the next + * generation and keep the top half perfomers + * + * This assumes natural_select_cutoff is exactly half of num_children + * and num_children is a multable of 4. + */ +static void genetic_ns_top_parents(phenotype_t * pt) +{ + unsigned long i,j,k = 0; + unsigned long num_children = pt->num_children; + unsigned long cutoff = num_children - pt->natural_selection_cutoff; + + for (i = cutoff, j = num_children - 1; i < j; i++, j--, k += 2) { + /* create child A */ + pt->ops->combine_genes(pt->child_ranking[i], + pt->child_ranking[j], + pt->child_ranking[k]); + + /* create child B */ + pt->ops->combine_genes(pt->child_ranking[i], + pt->child_ranking[j], + pt->child_ranking[k+1]); + } +} + + +/* This natural selection routine just has top parents populating + bottom performers. */ +static void genetic_ns_award_top_parents(phenotype_t * pt) +{ + unsigned long i; + unsigned long num_children = pt->num_children; + unsigned long cutoff = num_children - pt->natural_selection_cutoff; + + for (i = 0; i < cutoff; i += 2) { + pt->ops->combine_genes(pt->child_ranking[num_children - 1], + pt->child_ranking[num_children - 2], + pt->child_ranking[i]); + + pt->ops->combine_genes(pt->child_ranking[num_children - 1], + pt->child_ranking[num_children - 2], + pt->child_ranking[i+1]); + } +} + +static inline void genetic_swap(genetic_child_t ** a, genetic_child_t ** b) +{ + genetic_child_t * tmp = *a; + + *a = *b; + *b = tmp; +} + +/* bubble sort */ +/* XXX change this to quick sort */ +static void genetic_split_performers(phenotype_t * pt) +{ + int i, j; + + for (i = pt->num_children; i > 1; i--) + for (j = 0; j < i - 1; j++) + if (pt->child_ranking[j]->fitness > pt->child_ranking[j+1]->fitness) + genetic_swap(&pt->child_ranking[j], &pt->child_ranking[j+1]); +} + +static void genetic_mutate(phenotype_t * pt) +{ + long child_entry = -1; + int i; + + if (!pt->num_genes) + return; + + for (i = 0; i < pt->num_mutations; i++) { + get_random_bytes(&child_entry, sizeof(child_entry)); + child_entry = child_entry % pt->num_children; + + pt->ops->mutate_child(pt->child_ranking[child_entry]); + } +} + +/* XXX This will either aid in handling new workloads, or send us on a + downward spiral */ +static void genetic_shift_mutation_rate(phenotype_t * pt, long long prev_gen_avg_fitness, long long avg_fitness) +{ + + long long low_bound; + long long high_bound; + long dummy; + + if (mutation_rate_change && pt->genetic->generation_number > 1) { + + if (pt->ops->shift_mutation_rate) { + pt->ops->shift_mutation_rate(pt); + } else { + + low_bound = avg_fitness * 90; + divll(&low_bound, 100, &dummy); + + high_bound = avg_fitness * 110; + divll(&high_bound, 100, &dummy); + + if (high_bound > prev_gen_avg_fitness) + pt->mutation_rate -= mutation_rate_change; + else if (low_bound < prev_gen_avg_fitness) + pt->mutation_rate += mutation_rate_change; + + if (pt->mutation_rate > GENETIC_MAX_MUTATION_RATE) + pt->mutation_rate = GENETIC_MAX_MUTATION_RATE; + else if (pt->mutation_rate < GENETIC_MIN_MUTATION_RATE) + pt->mutation_rate = GENETIC_MIN_MUTATION_RATE; + } + } +} + +void genetic_general_shift_mutation_rate(phenotype_t * in_pt) +{ + struct list_head * p; + phenotype_t * pt; + int count = 0; + long rate = 0; + + list_for_each(p, &in_pt->genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + + if (in_pt->uid & pt->uid && in_pt->uid != pt->uid) { + rate += pt->mutation_rate; + count++; + } + } + + /* If we are a general phenotype that is made up of other + phenotypes then we take the average */ + if (count) + in_pt->mutation_rate = (rate / count); + else + in_pt->mutation_rate = mutation_rate_change; +} + +static void genetic_calc_stats(phenotype_t * in_pt) +{ + struct list_head * p; + phenotype_t * pt; + long long total_fitness = 0; + long long prev_gen_avg_fitness = in_pt->last_gen_avg_fitness; + long long tmp_fitness; + long dummy; + int i = 0; +#ifdef CONFIG_FINGERPRINTING + int fp = in_pt->genetic->fingerprinting; + int numerical_fp; +#endif + + /* On a general phenotype, need to look at other metrics since + * the fitness is normalized. It always average the same. It + * assumes that this phenotype is registered last. + */ + if (in_pt->ops->calc_post_fitness) { + +#ifdef CONFIG_FINGERPRINTING + if (fp) + numerical_fp = create_fingerprint(in_pt->fp); + + /* do we want this???? */ + if ((fp && (in_pt->last_fingerprint == numerical_fp)) || !fp) { +#else + if (1) { +#endif + +#ifdef GENETIC_DEBUG_VERBOSE + printk(KERN_INFO "genetic_calc_stats() for %s\n", in_pt->name); +#endif + list_for_each(p, &in_pt->genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + + /* for each child */ + if (in_pt->uid & pt->uid && in_pt->uid != pt->uid) { + if (pt->avg_fitness) { + /* measure how far percentage-wise that we are from the top */ + pt->from_top = (pt->last_gen_avg_fitness - pt->avg_fitness) * 100; +#ifdef GENETIC_DEBUG_VERBOSE + printk(" name: %s from_top: %lld avg_fitness: %lld\n", pt->name, pt->from_top, pt->avg_fitness); +#endif + divll(&pt->from_top, (pt->avg_fitness > 0) ? pt->avg_fitness : -pt->avg_fitness, &dummy); + + total_fitness += pt->from_top; +#ifdef GENETIC_DEBUG_VERBOSE + printk(" total_fitness: %lld\n", total_fitness); +#endif + } + } + + i++; + + } + + } else { + /* XXX horrible horrible hack...but + * testing viability */ + total_fitness = 0; + i = 1; + } + + BUG_ON(!i); + + in_pt->last_gen_avg_fitness = total_fitness; + divll(&in_pt->last_gen_avg_fitness, i, &dummy); + +#ifdef GENETIC_DEBUG_VERBOSE + printk(" in_pt->last_gent_avg_fitness: %lld\n", in_pt->last_gen_avg_fitness); +#endif + } else { + /* calculate the avg fitness for this generation and avg fitness + * so far */ + for (i = 0; i < in_pt->num_children; i++) + total_fitness += in_pt->child_ranking[i]->fitness; + + in_pt->last_gen_avg_fitness = total_fitness >> ilog2(in_pt->num_children); + } + + /* Mutation rate calibration */ + genetic_shift_mutation_rate(in_pt, prev_gen_avg_fitness, + in_pt->last_gen_avg_fitness); + + in_pt->num_mutations = ((in_pt->num_children * in_pt->num_genes) * in_pt->mutation_rate) / 100; + + /* calc new avg fitness */ + tmp_fitness = in_pt->last_gen_avg_fitness - in_pt->avg_fitness; + divll(&tmp_fitness, in_pt->genetic->generation_number, &dummy); + in_pt->avg_fitness += tmp_fitness; + + in_pt->fitness_history[in_pt->fitness_history_index++ & GENETIC_HISTORY_MASK] = + in_pt->last_gen_avg_fitness; + +#ifdef GENETIC_DEBUG_VERBOSE + printk("finished genetic_calc_stats()\n"); +#endif +} + + +void genetic_new_generation(genetic_t * genetic) +{ + struct list_head * tmp; + + struct list_head * p; + phenotype_t * pt; + + list_for_each(p, &genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + + /* Check to see if need to recalibrate fitness to take + other phenotypes' rankings into account. This + should be ran after all phenotypes that have input + have been ran. */ + if (pt->ops->calc_post_fitness) + pt->ops->calc_post_fitness(pt); + + dump_children(pt); + + /* figure out top performers */ + genetic_split_performers(pt); + + /* calc stats */ + genetic_calc_stats(pt); + + dump_children(pt); + + /* make some new children */ + if (pt->num_genes) + pt->natural_selection(pt); + + dump_children(pt); + +#ifdef CONFIG_FINGERPRINTING + if (pt->ops->get_fingerprint) { + + pt->ops->get_fingerprint(pt); + reset_fp_snapshot(pt->fp_ss); + + /* See if this generation was a top performer + * for the current workload. + * Do this after natural selection to get rid + * of the bad apples + */ + update_top_performers(pt); + + /* We know the workload, lets put some known + good genes back in */ + reintroduce_genes(pt); + + pt->last_fingerprint = create_fingerprint(pt->fp); + } + + dump_children(pt); +#endif + + /* mutate a couple of the next generation */ + genetic_mutate(pt); + + dump_children(pt); + + /* Move the new children still sitting in the finished queue to + the run queue */ + tmp = pt->run_queue; + pt->run_queue = pt->finished_queue; + pt->finished_queue = tmp; + + pt->child_number = 0; +#if GENETIC_DEBUG + pt->debug_index = 0; +#endif + + } + + genetic->child_number = 0; + genetic->generation_number++; + +} + +/** + * genetic_generic_random_mutate_gene - mutate child's gene to value in range + * @child: child whose gene we are mutating + * @gene_num: gene index from gene_param to mutate; gene must be unsigned long + * + * Mutate a gene picking a random value within the gene range that was + * specified in @child->gene_param. + */ +void genetic_generic_random_mutate_gene(genetic_child_t * child, + unsigned long gene_num) +{ + unsigned long *genes = (unsigned long *)child->genes; + unsigned long min = child->gene_param[gene_num].min; + unsigned long max = child->gene_param[gene_num].max; + unsigned long gene_value; + unsigned long range = max - min + 1; + + /* create a mutation value */ + get_random_bytes(&gene_value, sizeof(gene_value)); + + gene_value = gene_value % range; + + genes[gene_num] = min + gene_value; +} + +/** + * genetic_generic_iterative_mutate_gene + * @child: child whose gene we are mutating + * @gene_num: gene index from gene_param to mutate; gene must be unsigned long + */ +void genetic_generic_iterative_mutate_gene(genetic_child_t * child, + unsigned long gene_num) +{ + unsigned long *genes = (unsigned long *)child->genes; + unsigned long min = child->gene_param[gene_num].min; + unsigned long max = child->gene_param[gene_num].max; + long change; + unsigned long old_value = genes[gene_num]; + unsigned long new_value; + unsigned long range = max - min + 1; + + /* If under 5, random might work better */ + if (range < 5) + return genetic_generic_random_mutate_gene(child, gene_num); + + /* get the % of change */ + get_random_bytes(&change, sizeof(change)); + + change = change % GENETIC_ITERATIVE_MUTATION_RANGE; + + new_value = ((long)(change * range) / (long)100) + old_value; + + if (new_value > max) + new_value = max; + else if (new_value < min) + new_value = min; + + genes[gene_num] = new_value; +} + +/** + * genetic_generic_mutate_child - mutate random gene in child + * @child: child whose gene we are mutating. + * + * Select a random gene and mutate it either using either the mutate_gene + * callback specified in '&struct gene_param' OR if that is NULL then use + * 'genetic_generic_random_mutate_gene()' + */ +void genetic_generic_mutate_child(genetic_child_t * child) +{ + long gene_num = -1; + + /* pick a random gene */ + get_random_bytes(&gene_num, sizeof(gene_num)); + + if (gene_num < 0) + gene_num = -gene_num; + + gene_num = gene_num % child->num_genes; + + if (child->gene_param[gene_num].mutate_gene) + child->gene_param[gene_num].mutate_gene(child, gene_num); + else + genetic_generic_random_mutate_gene(child, gene_num); +} + +/** + * genetic_generic_mutate_child - set all genes to their initial value + */ +void genetic_create_child_defaults(genetic_child_t * child) +{ + int i; + unsigned long * genes = child->genes; + + for (i = 0; i < child->num_genes; i++) { + genes[i] = child->gene_param[i].initial; + } +} + +void genetic_create_child_spread(genetic_child_t * child, + unsigned long num_children) +{ + int i; + unsigned long range; + int range_incr; + int child_num = child->id; + long num_genes = child->num_genes; + unsigned long * genes = child->genes; + + for (i = 0; i < num_genes; i++) { + range = child->gene_param[i].max - child->gene_param[i].min + 1; + range_incr = range / num_children; + if (range_incr) + genes[i] = child->gene_param[i].min + + (range_incr * child_num); + else + genes[i] = child->gene_param[i].min + + (child_num / (num_children / range)); + } + +} + +#if 0 +/* Randomly pick which parent to use for each gene to create a child */ +void genetic_generic_combine_genes(genetic_child_t * parent_a, + genetic_child_t * parent_b, + genetic_child_t * child) +{ + unsigned long * genes_a = (unsigned long *)parent_a->genes; + unsigned long * genes_b = (unsigned long *)parent_b->genes; + unsigned long * child_genes = (unsigned long *)child->genes; + + /* Assume parent_a and parent_b have same num_genes */ + unsigned long num_genes = parent_a->num_genes; + int parent_selector; + int i; + + get_random_bytes(&parent_selector, sizeof(parent_selector)); + + if ((sizeof(parent_selector) * 8) < num_genes) + BUG(); + + for (i = 0; i < num_genes; i++) { + /* Look at each bit to determine which parent to use */ + if (parent_selector & 1) { + child_genes[i] = genes_a[i]; + } else { + child_genes[i] = genes_b[i]; + } + parent_selector >>= 1; + } +} +#else + +/** + * genetic_generic_combine_genes - create child using comb. of parent's genes + * @parent_a: gene doner + * @parent_b: gene doner + * @child: genes will be modified using a combination of a and b + */ +void genetic_generic_combine_genes(genetic_child_t * parent_a, + genetic_child_t * parent_b, + genetic_child_t * child) +{ + unsigned long * genes_a = (unsigned long *)parent_a->genes; + unsigned long * genes_b = (unsigned long *)parent_b->genes; + unsigned long * child_genes = (unsigned long *)child->genes; + + /* Assume parent_a and parent_b have same num_genes */ + unsigned long num_genes = parent_a->num_genes; + int percentage; + int i; + + for (i = 0; i < num_genes; i++) { + get_random_bytes(&percentage, sizeof(percentage)); + + /* Get percentage */ + percentage = percentage % 100; + + if (percentage < 0) + percentage = -percentage; + + /* Give child x% of parent A's genes value, plus + 100-x% of parent B's genes value */ + child_genes[i] = (((genes_a[i]+1) * percentage) + + (genes_b[i] * (100 - percentage))) / 100; + } +} +#endif + +static int __init genetic_boot_setup(char *str) +{ + if (strcmp(str, "on") == 0) + genetic_lib_enabled = 1; + else if (strcmp(str, "off") == 0) + genetic_lib_enabled = 0; + + return 1; +} + + +static int __init genetic_mutation_rate_change_setup(char *str) +{ + int i; + + if (get_option(&str,&i)) { + + if (i > GENETIC_MAX_MUTATION_RATE) + i = GENETIC_MAX_MUTATION_RATE; + else if (i < 0) + i = 0; + + mutation_rate_change = i; + } + + return 1; + +} +__setup("genetic=", genetic_boot_setup); +__setup("genetic_mutate_rate=", genetic_mutation_rate_change_setup); Index: linux-2.6.23/lib/genetic-debug.c =================================================================== --- /dev/null +++ linux-2.6.23/lib/genetic-debug.c @@ -0,0 +1,153 @@ +/* + * Genetic Algorithm Library Debugging Routines + * + * (C) Copyright 2006 IBM + * (C) Copyright 2006 Brandon Philips + * + * This program is free software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as published + * by the Free Software Foundation. + */ + +#include +#include +#include + +struct dentry * genetic_tree_root = NULL; + +/** + * genetic_stat_show - generates /debug/genetic//stats + */ +static int genetic_stat_show(struct seq_file *s, void *unused) +{ + genetic_t * genetic = (genetic_t *)s->private; + + seq_printf(s, "name: %s\n", genetic->name); + seq_printf(s, "generation_number: %ld\n", genetic->generation_number); + seq_printf(s, "num_children: %ld\n", genetic->num_children); + seq_printf(s, "child_life_time: %ld\n", genetic->child_life_time); + seq_printf(s, "child_number: %ld\n", genetic->child_number); + + return 0; +} + +static int genetic_stat_open(struct inode *inode, struct file *file) +{ + return single_open(file, genetic_stat_show, inode->i_private); +} + +static struct file_operations genetic_stat_operations = { + .open = genetic_stat_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + +/** + * genetic_phenotype_average_show - /debug/genetic//phenotype_average + */ +static int genetic_phenotype_average_show(struct seq_file *s, void *unused) +{ + genetic_t * genetic = (genetic_t *)s->private; + struct list_head * p; + phenotype_t * pt; + + list_for_each(p, &genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + seq_printf(s, "%s: %lld\n", pt->name, pt->avg_fitness); + } + + return 0; +} + +static int genetic_phenotype_average_open(struct inode *inode, struct file *file) +{ + return single_open(file, genetic_phenotype_average_show, inode->i_private); +} + +static struct file_operations genetic_phenotype_average_operations = { + .open = genetic_phenotype_average_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + + +/** + * genetic_genes_show - /debug/genetic//gene + */ +int genetic_generic_gene_show(struct seq_file *s, void *unused) +{ + int i; + phenotype_t * pt = (phenotype_t *)s->private; + + genetic_child_t * child = list_entry(pt->run_queue->next, + genetic_child_t, list); + + unsigned long * genes = (unsigned long *)child->genes; + + for (i = 0; i < pt->num_genes; i++) + seq_printf(s, "%s: %lu\n", child->gene_param[i].name, genes[i]); + + return 0; +} + +static int genetic_generic_gene_open(struct inode *inode, struct file *file) +{ + phenotype_t * pt = (phenotype_t *)inode->i_private; + + return single_open(file, pt->ops->gene_show, inode->i_private); +} + +static struct file_operations genetic_gene_operations = { + .open = genetic_generic_gene_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + + +static struct dentry *genetic_create_tree(const char *name, struct dentry *parent) +{ + struct dentry *dir = NULL; + + if (!genetic_tree_root) { + genetic_tree_root = debugfs_create_dir("genetic", NULL); + if (!genetic_tree_root) + goto err; + } + + if (!parent) parent = genetic_tree_root; + + dir = debugfs_create_dir(name, parent); + +err: + return dir; +} + +#if GENETIC_DEBUG +/* Stores attributes into an array in the following format + * child_num fitness gene[0] gene[1] .... gene[num_genes-1] + * Add +1 to GENETIC_NUM_DEBUG_POINTS if add another dump_children + * call + */ +void dump_children(phenotype_t * pt) +{ + int i, j; + long * genes; + unsigned long debug_size = pt->debug_size; + + for (i = 0; i < pt->num_children; i++) { + pt->debug_history[pt->debug_index++ % debug_size] = pt->child_ranking[i]->id; + pt->debug_history[pt->debug_index++ % debug_size] = pt->child_ranking[i]->fitness; + + genes = (long *)pt->child_ranking[i]->genes; + + for (j = 0; j < pt->child_ranking[i]->num_genes; j++) { + pt->debug_history[pt->debug_index++ % debug_size] = genes[j]; + } + } +} +#else +void dump_children(phenotype_t * pt) {} +#endif /* GENETIC_DEBUG */ Index: linux-2.6.23/lib/fingerprinting.c =================================================================== --- /dev/null +++ linux-2.6.23/lib/fingerprinting.c @@ -0,0 +1,291 @@ +static int create_fingerprint(struct fingerprint * fp) +{ + int numerical_fp = 0; + + numerical_fp |= fp->type; + numerical_fp <<= 1; + + numerical_fp |= fp->pattern; + numerical_fp <<= 1; + + numerical_fp |= fp->size; + + return numerical_fp; +} + +static long long get_top_fitness(phenotype_t * pt, struct fingerprint * fp) +{ + return pt->top_fitness[fp->type][fp->pattern][fp->size]; +} + + +static int top_fitness_open(struct inode *inode, struct file *file) +{ + phenotype_t * pt = (phenotype_t *)inode->i_private; + + return single_open(file, pt->ops->top_fitness_show, inode->i_private); +} + +static struct file_operations top_fitness_ops = { + .open = top_fitness_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + +static int snapshot_open(struct inode *inode, struct file *file) +{ + phenotype_t * pt = (phenotype_t *)inode->i_private; + + return single_open(file, pt->ops->snapshot_show, inode->i_private); +} + +static struct file_operations snapshot_ops = { + .open = snapshot_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + +static int state_open(struct inode *inode, struct file *file) +{ + phenotype_t * pt = (phenotype_t *)inode->i_private; + + return single_open(file, pt->ops->state_show, inode->i_private); +} + +static struct file_operations state_ops = { + .open = state_open, + .read = seq_read, + .llseek = seq_lseek, + .release = single_release, +}; + + +int genetic_init_fingerprinting(phenotype_t * pt) +{ + int i, j, k; + struct genetic_ops * ops = pt->ops; + int num_genes = pt->num_genes; + + if (num_genes) { + + pt->fp = (struct fingerprint *)kmalloc( + sizeof(struct fingerprint), GFP_KERNEL); + + if (!pt->fp) { + printk(KERN_ERR "genetic_register_phenotype: not enough" + "memory\n"); + return -ENOMEM; + } + + reset_fp(pt->fp); + + pt->fp_ss = (struct fp_snapshot *)kmalloc( + sizeof(struct fp_snapshot), GFP_KERNEL); + + if (!pt->fp_ss) { + printk(KERN_ERR "genetic_register_phenotype: not enough" + "memory\n"); + return -ENOMEM; + } + + reset_fp_snapshot(pt->fp_ss); + + pt->top_child = (unsigned long ***)kmalloc( + sizeof(unsigned long ***) * 2, GFP_KERNEL); + + if (!pt->top_child) { + printk(KERN_ERR "genetic_register_phenotype: not enough" + "memory\n"); + return -ENOMEM; + } + + for (i = 0; i < 2; i++) { + pt->top_child[i] = (unsigned long **)kmalloc( + sizeof(unsigned long **) * 2, + GFP_KERNEL); + + if (!pt->top_child[i]) { + printk(KERN_ERR "genetic_register_phenotype:\ + not enough memory\n"); + return -ENOMEM; + } + + for (j = 0; j < 2; j++) { + pt->top_child[i][j] = (unsigned long *)kmalloc( + sizeof(unsigned long *) * 2, + GFP_KERNEL); + + if (!pt->top_child[i][j]) { + printk(KERN_ERR "genetic_register_phenotype: not enough memory\n"); + return -ENOMEM; + } + + for (k = 0; k < 2; k++) { + pt->top_child[i][j][k] = (unsigned long)ops->create_top_genes(pt); + if (!pt->top_child[i][j][k]) + return -ENOMEM; + } + } + } + } /* if (num_genes) */ + + pt->top_fitness = (long long ***)kmalloc(sizeof(long long ***) * 2, GFP_KERNEL); + if (!pt->top_fitness) { + printk(KERN_ERR "genetic_register_phenotype: not enough" + "memory\n"); + return -ENOMEM; + } + + for (i = 0; i < 2; i++) { + pt->top_fitness[i] = (long long **)kmalloc(sizeof(long long **) * 2, GFP_KERNEL); + if (!pt->top_fitness[i]) { + printk(KERN_ERR "genetic_register_phenotype: not" + "enough memory\n"); + return -ENOMEM; + } + + for (j = 0; j < 2; j++) { + pt->top_fitness[i][j] = (long long *)kmalloc( + sizeof(long long *) * 2, + GFP_KERNEL); + + if (!pt->top_fitness[i][j]) { + printk(KERN_ERR "genetic_register_phenotype: " + "not enough memory\n"); + return -ENOMEM; + } + + for (k = 0; k < 2; k++) { + pt->top_fitness[i][j][k] = 0; + } + } + } + + pt->last_fingerprint = 0; + + if (pt->genetic->fingerprinting_dir) { + pt->fp_dir = genetic_create_tree(pt->name, + pt->genetic->fingerprinting_dir); + + if (ops->top_fitness_show) + debugfs_create_file("top_fitness", S_IFREG|S_IRUGO, + pt->fp_dir, pt, &top_fitness_ops); + + if (ops->snapshot_show) + debugfs_create_file("snapshot", S_IFREG|S_IRUGO, + pt->fp_dir, pt, &snapshot_ops); + + if (ops->state_show) + debugfs_create_file("state", S_IFREG|S_IRUGO, + pt->fp_dir, pt, &state_ops); + } + + return 0; +} + +static void decay_fitness(phenotype_t * pt, struct fingerprint * fp) +{ + long long fitness; + long dummy; + + fitness = get_top_fitness(pt, fp); + + /* reduce the fitness to eventually get new genes in */ + fitness *= FP_DECAY; + divll(&fitness, 100, &dummy); + + pt->top_fitness[fp->type][fp->pattern][fp->size] = fitness; +} + +static void update_phenotype_top_performer(phenotype_t * pt, struct fingerprint * fp) +{ + long long top_fitness; + unsigned long * genes; + long long * avg_genes; + long dummy; + int i, j; + + + /* Decay the top fitness so not to have a fluke and have a + * high set which are less than optimal. So decay the top + * fitness so eventually these genes are phased out. + */ + decay_fitness(pt, fp); + + top_fitness = get_top_fitness(pt, fp); + + if (pt->last_gen_avg_fitness >= top_fitness) { + + pt->top_fitness[fp->type][fp->pattern][fp->size] = pt->last_gen_avg_fitness; + + /* We don't need to track this if there's no genes! */ + if (!pt->num_genes) + return; + + avg_genes = (long long *)kmalloc(sizeof(long long) * pt->num_genes, GFP_KERNEL); + if (!avg_genes) { + printk(KERN_ERR "update_top_performers: unable to alloc space\n"); + return; + } + + memset(avg_genes, 0, sizeof(long long) * pt->num_genes); + + for (i = 0; i < pt->num_genes; i++) { + for (j = 0; j < pt->num_children; j++) { + genes = pt->child_ranking[j]->genes; + avg_genes[i] += genes[i]; + } + } + + for (j = 0; j < pt->num_genes; j++) + divll(&avg_genes[j], pt->num_children, &dummy); + + genes = (unsigned long *)pt->top_child[fp->type][fp->pattern][fp->size]; + for (j = 0; j < pt->num_genes; j++) + genes[j] = avg_genes[j]; + + kfree(avg_genes); + } +} + +static void update_top_performers(phenotype_t * master) +{ + phenotype_t * pt; + struct list_head * p; + + list_for_each(p, &master->genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + + if (master->uid & pt->uid && master->uid != pt->uid) { + update_phenotype_top_performer(pt, master->fp); + } + } + update_phenotype_top_performer(master, master->fp); +} + +static void reintroduce_genes(phenotype_t * master) +{ + struct fingerprint * fp = master->fp; + phenotype_t * pt; + unsigned long * top_genes; + unsigned long * genes; + struct list_head * p; + int i; + + list_for_each(p, &master->genetic->phenotype) { + pt = list_entry(p, phenotype_t, phenotype); + + if (pt->num_genes) { + + /* Do this more intelligently, so can have n-points on + the fingerprint */ + /* just take the first one */ + top_genes = (unsigned long *)pt->top_child[fp->type][fp->pattern][fp->size]; + genes = pt->child_ranking[0]->genes; + for (i = 0; i < pt->num_children; i++) + genes[i] = top_genes[i]; + } + } +} Index: linux-2.6.23/include/linux/fingerprinting.h =================================================================== --- /dev/null +++ linux-2.6.23/include/linux/fingerprinting.h @@ -0,0 +1,125 @@ +#ifndef __LINUX_FINGERPRINTING_H +#define __LINUX_FINGERPRINTING_H + +/* + * include/linux/fingerprinting.h + * + * Jake Moilanen + * Copyright (C) 2006 IBM + * + * I/O Workload Fingerprinting + * + * This program is free software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as published + * by the Free Software Foundation. +*/ + +#include +#include + +#define FP_TYPE_READ 0 +#define FP_TYPE_WRITE 1 +#define FP_PATTERN_SEQ 0 +#define FP_PATTERN_RAND 1 +#define FP_SIZE_SMALL 0 +#define FP_SIZE_LARGE 1 +#define FP_NUM_POINTS (2 * 2 * 2) + +struct fingerprint { + __u8 type; + __u8 pattern; + __u8 size; +}; + +struct fp_snapshot { + /* type */ + unsigned long reads; + unsigned long writes; + /* pattern */ + unsigned long head_pos; + unsigned long avg_dist; + /* size */ + unsigned long avg_size; +}; + +/* Number of reads/writes before classified as read */ +#define FP_CLASS_READ_WRITE_RATIO 2 + +/* Number of sectors before pattern is random */ +#define FP_CLASS_PATTERN_RAND 25 + +/* Number of sectors before size is large */ +#define FP_CLASS_SIZE_LARGE 8 + +extern void update_fp_snapshot(struct bio * bio); +extern void calc_fp(struct fingerprint * fp, struct fp_snapshot * fp_ss); +extern void reset_fp_snapshot(struct fp_snapshot * ss); +extern void reset_fp(struct fingerprint * fp); +extern void consolidate_fp_snapshot(struct fp_snapshot * master, struct fp_snapshot * instance); +extern int fingerprint_state_show(struct seq_file *s, void *unused); +extern int fingerprint_snapshot_show(struct seq_file *s, void *unused); +extern int fingerprint_top_fitness_show(struct seq_file *s, void *unused); + +/* XXX do this more intelligently */ +#ifndef DIVLL_OP +#define DIVLL_OP +#if BITS_PER_LONG >= 64 + +static inline void divll(long long *n, long div, long *rem) +{ + *rem = *n % div; + *n /= div; +} + +#else + +static inline void divl(int32_t high, int32_t low, + int32_t div, + int32_t *q, int32_t *r) +{ + int64_t n = (u_int64_t)high << 32 | low; + int64_t d = (u_int64_t)div << 31; + int32_t q1 = 0; + int c = 32; + while (n > 0xffffffff) { + q1 <<= 1; + if (n >= d) { + n -= d; + q1 |= 1; + } + d >>= 1; + c--; + } + q1 <<= c; + if (n) { + low = n; + *q = q1 | (low / div); + *r = low % div; + } else { + *r = 0; + *q = q1; + } + return; +} + +static inline void divll(long long *n, long div, long *rem) +{ + int32_t low, high; + low = *n & 0xffffffff; + high = *n >> 32; + if (high) { + int32_t high1 = high % div; + int32_t low1 = low; + high /= div; + divl(high1, low1, div, &low, (int32_t *)rem); + *n = (int64_t)high << 32 | low; + } else { + *n = low / div; + *rem = low % div; + } +} +#endif + +#endif /* #ifndef divll */ + +#endif /* __LINUX_FINGERPRINTINT_H */ Index: linux-2.6.23/block/fingerprinting.c =================================================================== --- /dev/null +++ linux-2.6.23/block/fingerprinting.c @@ -0,0 +1,205 @@ +/* + * block/fingerprinting.c + * + * Jake Moilanen + * Copyright (C) 2006 IBM + * + * I/O Workload Fingerprinting + * + * This program is free software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as published + * by the Free Software Foundation. +*/ +/* TODOS: + * - Abstract so no so IO specific + * - Abstract types + */ + +#include +#include +#include +#include + +int fingerprint_state_show(struct seq_file *s, void *unused) +{ + phenotype_t * pt = (phenotype_t *)s->private; + struct fingerprint * fp = pt->fp; + + if (fp->type == FP_TYPE_READ) + seq_printf(s, "read\t(%d)\n", FP_TYPE_READ); + else + seq_printf(s, "write\t(%d)\n", FP_TYPE_WRITE); + + if (fp->pattern == FP_PATTERN_SEQ) + seq_printf(s, "sequential\t(%d)\n", FP_PATTERN_SEQ); + else + seq_printf(s, "random\t(%d)\n", FP_PATTERN_RAND); + + if (fp->size == FP_SIZE_SMALL) + seq_printf(s, "small\t(%d)\n", FP_SIZE_SMALL); + else + seq_printf(s, "large\t(%d)\n", FP_SIZE_LARGE); + + return 0; +} + +int fingerprint_snapshot_show(struct seq_file *s, void *unused) +{ + phenotype_t * pt = (phenotype_t *)s->private; + struct fp_snapshot * ss = pt->fp_ss; + + seq_printf(s, "read: %ld\n", ss->reads); + seq_printf(s, "write: %ld\n", ss->writes); + + seq_printf(s, "avg_dist: %ld\n", ss->avg_dist); + seq_printf(s, "avg_size: %ld\n", ss->avg_size); + + return 0; +} + + +int fingerprint_top_fitness_show(struct seq_file *s, void *unused) +{ + int i, j, k; + phenotype_t * pt = (phenotype_t *)s->private; + + for (i = 0; i < 2; i++) + for (j = 0; j < 2; j++) + for (k = 0; k < 2; k++) + seq_printf(s, "top_fitness[%d][%d][%d]: %lld\n", + i, j, k, pt->top_fitness[i][j][k]); + + return 0; +} + + +/* This assumes that address matches up w/ head_pos */ +static void update_avg_dist(struct fp_snapshot * ss, long head_pos) +{ + long long tmp_dist; + unsigned long total_ops = ss->reads + ss->writes; + long dummy; + + /* set it the first time through */ + if (!ss->head_pos) { + ss->head_pos = head_pos; + return; + } + tmp_dist = ss->head_pos - head_pos; + if (tmp_dist < 0) + tmp_dist = -tmp_dist; + + tmp_dist = tmp_dist - ss->avg_dist; + + divll(&tmp_dist, total_ops, &dummy); + ss->avg_dist += tmp_dist; + + ss->head_pos = head_pos; + +} + +static void update_avg_size(struct fp_snapshot * ss, unsigned long size) +{ + unsigned long total_ops = ss->reads + ss->writes; + long long tmp_size; + long dummy; + + tmp_size = size - ss->avg_size; + divll(&tmp_size, total_ops, &dummy); + ss->avg_size += tmp_size; +// ss->avg_size += (size - ss->avg_size) / total_ops; +} + +void update_fp_snapshot(struct bio * bio) +{ + struct fp_snapshot * ss = bio->bi_bdev->bd_disk->fp_ss; + + /* update type */ + if (bio_data_dir(bio) == READ) + ss->reads++; + else + ss->writes++; + + /* update pattern */ +// update_avg_dist(ss, bio_to_phys(bio)); + update_avg_dist(ss, bio->bi_sector); + + /* update size */ +// update_avg_size(ss, bio_iovec(bio)->bv_len); + update_avg_size(ss, bio_sectors(bio)); + +} + +/* Use this when there's multiple disks, and need to consolidate to a + * system wide fingerprint + */ +void consolidate_fp_snapshot(struct fp_snapshot * master, struct fp_snapshot * instance) +{ + unsigned long total_ops; + long dummy; + long long total_dist; + long long total_size; + + BUG_ON(!master); + BUG_ON(!instance); + + total_dist = master->avg_dist * (master->reads + master->writes); + total_size = master->avg_size * (master->reads + master->writes); + + /* update operations */ + master->reads += instance->reads; + master->writes += instance->writes; + total_ops = master->reads + master->writes; + + /* update distance */ + total_dist += (instance->avg_dist * (instance->reads + instance->writes)); + if (total_ops) { + divll(&total_dist, total_ops, &dummy); + master->avg_dist = total_dist; + } else + master->avg_dist = 0; + + /* update size */ + total_size += (instance->avg_size * (instance->reads + instance->writes)); + if (total_ops) { + divll(&total_size, total_ops, &dummy); + master->avg_size = total_size; + } else + master->avg_size = 0; +} + +void reset_fp_snapshot(struct fp_snapshot * ss) +{ + memset(ss, 0, sizeof(struct fp_snapshot)); +} + +void reset_fp(struct fingerprint * fp) +{ + memset(fp, 0, sizeof(struct fingerprint)); +} + +//void calc_fp(struct fingerprint * fp, struct fp_snapshot * fp_ss, struct block_device * dev) +void calc_fp(struct fingerprint * fp, struct fp_snapshot * fp_ss) +{ + /* type */ + if (fp_ss->reads > (fp_ss->writes * FP_CLASS_READ_WRITE_RATIO)) + fp->type = FP_TYPE_READ; + else + fp->type = FP_TYPE_WRITE; + + /* pattern */ +// if (fp_ss->avg_dist >= (block_size(dev) * FP_CLASS_PATTERN_RAND)) + if (fp_ss->avg_dist >= (512 * FP_CLASS_PATTERN_RAND)) + fp->pattern = FP_PATTERN_RAND; + else + fp->pattern = FP_PATTERN_SEQ; + + /* size */ + if (fp_ss->avg_size > FP_CLASS_SIZE_LARGE) + fp->size = FP_SIZE_LARGE; + else + fp->size = FP_SIZE_SMALL; +} + + + Index: linux-2.6.23/block/ll_rw_blk.c =================================================================== --- linux-2.6.23.orig/block/ll_rw_blk.c +++ linux-2.6.23/block/ll_rw_blk.c @@ -30,6 +30,7 @@ #include #include #include +#include /* * for max sense size @@ -2893,6 +2894,9 @@ static void init_request_from_bio(struct { req->cmd_type = REQ_TYPE_FS; +#ifdef CONFIG_FINGERPRINTING + update_fp_snapshot(bio); +#endif /* * inherit FAILFAST from bio (for read-ahead, and explicit FAILFAST) */ Index: linux-2.6.23/include/linux/genhd.h =================================================================== --- linux-2.6.23.orig/include/linux/genhd.h +++ linux-2.6.23/include/linux/genhd.h @@ -67,6 +67,7 @@ struct partition { #include #include #include +#include struct partition { unsigned char boot_ind; /* 0x80 - active */ @@ -91,6 +92,7 @@ struct gendisk { struct disk_stats dkstats; #endif struct work_struct async_notify; + struct fp_snapshot * fp_ss; }; /* Structure for sysfs attributes on block devices */ Index: linux-2.6.23/block/genhd.c =================================================================== --- linux-2.6.23.orig/block/genhd.c +++ linux-2.6.23/block/genhd.c @@ -445,6 +445,20 @@ static ssize_t disk_stats_read(struct ge jiffies_to_msecs(disk_stat_read(disk, io_ticks)), jiffies_to_msecs(disk_stat_read(disk, time_in_queue))); } +static ssize_t disk_fp_read(struct gendisk * disk, char *page) +{ + return sprintf(page, "reads: %llx\n" + "writes: %llx\n" + "head_pos: %llx\n" + "avg_dist: %llx\n" + "avg_size: %llx\n", + (unsigned long long)disk->fp_ss->reads, + (unsigned long long)disk->fp_ss->writes, + (unsigned long long)disk->fp_ss->head_pos, + (unsigned long long)disk->fp_ss->avg_dist, + (unsigned long long)disk->fp_ss->avg_size); +} + static struct disk_attribute disk_attr_uevent = { .attr = {.name = "uevent", .mode = S_IWUSR }, .store = disk_uevent_store @@ -473,6 +487,10 @@ static struct disk_attribute disk_attr_s .attr = {.name = "stat", .mode = S_IRUGO }, .show = disk_stats_read }; +static struct disk_attribute disk_attr_fp = { + .attr = {.name = "fp", .mode = S_IRUGO }, + .show = disk_fp_read +}; #ifdef CONFIG_FAIL_MAKE_REQUEST @@ -513,6 +531,7 @@ static struct attribute * default_attrs[ #ifdef CONFIG_FAIL_MAKE_REQUEST &disk_attr_fail.attr, #endif + &disk_attr_fp.attr, NULL, }; @@ -745,6 +764,10 @@ struct gendisk *alloc_disk_node(int mino INIT_WORK(&disk->async_notify, media_change_notify_thread); } + + disk->fp_ss = kmalloc(sizeof(struct fp_snapshot), GFP_KERNEL); + memset(disk->fp_ss, 0, sizeof(struct fp_snapshot)); + return disk; } Index: linux-2.6.23/block/Kconfig =================================================================== --- linux-2.6.23.orig/block/Kconfig +++ linux-2.6.23/block/Kconfig @@ -65,3 +65,9 @@ config BLK_DEV_BSG endif # BLOCK source block/Kconfig.iosched + +config FINGERPRINTING + bool "I/O Workload Fingerprinting" + help + Say Y here if you want workload data to be classified and + used to tune the I/O schedulers. Otherwise say N. \ No newline at end of file Index: linux-2.6.23/block/Makefile =================================================================== --- linux-2.6.23.orig/block/Makefile +++ linux-2.6.23/block/Makefile @@ -11,3 +11,6 @@ obj-$(CONFIG_IOSCHED_DEADLINE) += deadli obj-$(CONFIG_IOSCHED_CFQ) += cfq-iosched.o obj-$(CONFIG_BLK_DEV_IO_TRACE) += blktrace.o + + +obj-$(CONFIG_FINGERPRINTING) += fingerprinting.o