struct category *category_lookup_or_create(struct hash_table *categories, const char *name) {
struct category *c;
if(!name)
name = "default";
c = hash_table_lookup(categories, name);
if(c) return c;
c = calloc(1, sizeof(struct category));
c->name = xxstrdup(name);
c->fast_abort = -1;
c->total_tasks = 0;
c->first_allocation = NULL;
c->max_allocation = rmsummary_create(-1);
c->autolabel_resource = rmsummary_create(0);
c->max_resources_seen = rmsummary_create(-1);
c->cores_histogram = histogram_create(1);
c->wall_time_histogram = histogram_create(time_bucket_size);
c->cpu_time_histogram = histogram_create(time_bucket_size);
c->memory_histogram = histogram_create(memory_bucket_size);
c->swap_memory_histogram = histogram_create(memory_bucket_size);
c->virtual_memory_histogram = histogram_create(memory_bucket_size);
c->bytes_read_histogram = histogram_create(bytes_bucket_size);
c->bytes_written_histogram = histogram_create(bytes_bucket_size);
c->bytes_received_histogram = histogram_create(bytes_bucket_size);
c->bytes_sent_histogram = histogram_create(bytes_bucket_size);
c->bandwidth_histogram = histogram_create(bandwidth_bucket_size);
c->total_files_histogram = histogram_create(1);
c->disk_histogram = histogram_create(disk_bucket_size);
c->total_processes_histogram = histogram_create(1);
c->max_concurrent_processes_histogram = histogram_create(1);
c->time_peak_independece = 0;
c->steady_state = 0;
c->completions_since_last_reset = 0;
c->allocation_mode = CATEGORY_ALLOCATION_MODE_FIXED;
hash_table_insert(categories, name, c);
return c;
}
struct dag_node *dag_node_create(struct dag *d, int linenum)
{
struct dag_node *n;
n = malloc(sizeof(struct dag_node));
memset(n, 0, sizeof(struct dag_node));
n->d = d;
n->linenum = linenum;
n->state = DAG_NODE_STATE_WAITING;
n->nodeid = d->nodeid_counter++;
n->variables = hash_table_create(0, 0);
n->source_files = list_create(0);
n->target_files = list_create(0);
n->remote_names = itable_create(0);
n->remote_names_inv = hash_table_create(0, 0);
n->descendants = set_create(0);
n->ancestors = set_create(0);
n->ancestor_depth = -1;
n->resources_requested = rmsummary_create(-1);
n->resources_measured = NULL;
n->resource_request = CATEGORY_ALLOCATION_FIRST;
return n;
}
struct rmsummary *json_to_rmsummary(struct jx *j) {
if(!j || !jx_istype(j, JX_OBJECT))
return NULL;
struct rmsummary *s = rmsummary_create(-1);
struct jx_pair *head = j->u.pairs;
while(head) {
if(!jx_istype(head->key, JX_STRING))
continue;
char *key = head->key->u.string_value;
struct jx *value = head->value;
if(jx_istype(value, JX_STRING)) {
rmsummary_assign_char_field(s, key, value->u.string_value);
} else if(jx_istype(value, JX_INTEGER)) {
rmsummary_assign_int_field(s, key, value->u.integer_value);
} else if(jx_istype(value, JX_ARRAY)) {
int64_t number;
int status = json_number_of_array(value, key, &number);
if(status) {
rmsummary_assign_int_field(s, key, number);
}
}
head = head->next;
}
return s;
}
const struct rmsummary *category_dynamic_task_min_resources(struct category *c, struct rmsummary *user, category_allocation_t request) {
static struct rmsummary *internal = NULL;
const struct rmsummary *max = category_dynamic_task_max_resources(c, user, request);
if(internal) {
rmsummary_delete(internal);
}
internal = rmsummary_create(-1);
/* load seen values */
struct rmsummary *seen = c->max_resources_seen;
if(c->allocation_mode != CATEGORY_ALLOCATION_MODE_FIXED) {
internal->cores = seen->cores;
internal->memory = seen->memory;
internal->disk = seen->disk;
}
rmsummary_merge_override(internal, max);
return internal;
}
const struct rmsummary *category_dynamic_task_max_resources(struct category *c, struct rmsummary *user, category_allocation_t request) {
/* we keep an internal label so that the caller does not have to worry
* about memory leaks. */
static struct rmsummary *internal = NULL;
if(internal) {
rmsummary_delete(internal);
}
internal = rmsummary_create(-1);
struct rmsummary *max = c->max_allocation;
struct rmsummary *first = c->first_allocation;
struct rmsummary *seen = c->max_resources_seen;
if(c->steady_state && c->allocation_mode != CATEGORY_ALLOCATION_MODE_FIXED) {
internal->cores = seen->cores;
internal->memory = seen->memory;
internal->disk = seen->disk;
}
/* load max values */
rmsummary_merge_override(internal, max);
if(c->allocation_mode != CATEGORY_ALLOCATION_MODE_FIXED
&& request == CATEGORY_ALLOCATION_FIRST) {
rmsummary_merge_override(internal, first);
}
/* chip in user values */
rmsummary_merge_override(internal, user);
return internal;
}
void categories_initialize(struct hash_table *categories, struct rmsummary *top, const char *summaries_file) {
struct list *summaries = rmsummary_parse_file_multiple(summaries_file);
if(!summaries) {
fatal("Could not read '%s' file: %s\n", strerror(errno));
}
char *name;
struct category *c;
hash_table_firstkey(categories);
while(hash_table_nextkey(categories, &name, (void **) &c)) {
category_clear_histograms(c);
if(c->first_allocation) {
rmsummary_delete(c->first_allocation);
c->first_allocation = rmsummary_create(-1);
}
}
struct rmsummary *s;
list_first_item(summaries);
while((s = list_pop_head(summaries))) {
if(s->category) {
c = category_lookup_or_create(categories, s->category);
category_accumulate_summary(c, s, NULL);
}
rmsummary_delete(s);
}
hash_table_firstkey(categories);
while(hash_table_nextkey(categories, &name, (void **) &c)) {
category_update_first_allocation(c, NULL);
category_clear_histograms(c);
}
}
void rmDsummary_print(FILE *output, struct rmDsummary *so) {
struct rmsummary *s = rmsummary_create(-1);
s->command = xxstrdup(so->command);
if(so->category)
{
s->category = xxstrdup(so->category);
}
else if(so->command)
{
s->category = xxstrdup(so->command);
}
else
{
s->category = xxstrdup(DEFAULT_CATEGORY);
s->command = xxstrdup(DEFAULT_CATEGORY);
}
if(so->task_id) {
s->task_id = xxstrdup(so->task_id);
}
s->start = so->start;
s->end = so->end;
s->wall_time = so->wall_time;
to_internal(so, s, start, "us");
to_internal(so, s, end, "us");
to_internal(so, s, wall_time, "s");
to_internal(so, s, cpu_time, "s");
to_internal(so, s, cores, "cores");
to_internal(so, s, total_processes, "procs");
to_internal(so, s, max_concurrent_processes,"procs");
to_internal(so, s, memory, "MB");
to_internal(so, s, virtual_memory, "MB");
to_internal(so, s, swap_memory, "MB");
to_internal(so, s, bytes_read, "MB");
to_internal(so, s, bytes_written, "MB");
to_internal(so, s, bytes_received, "MB");
to_internal(so, s, bytes_sent, "MB");
to_internal(so, s, bandwidth, "Mbps");
to_internal(so, s, total_files, "files");
to_internal(so, s, disk, "MB");
rmsummary_print(output, s, /* pprint */ 1, /* extra fields */ 0);
rmsummary_delete(s);
return;
}
void category_specify_first_allocation_guess(struct category *c, const struct rmsummary *s) {
/* assume user knows what they are doing. */
c->steady_state = 1;
rmsummary_merge_max(c->max_resources_seen, s);
rmsummary_delete(c->first_allocation);
c->first_allocation = rmsummary_create(-1);
rmsummary_merge_max(c->first_allocation, s);
}
int main(int argc, char *argv[])
{
batch_queue_type_t batch_queue_type = BATCH_QUEUE_TYPE_UNKNOWN;
catalog_host = CATALOG_HOST;
catalog_port = CATALOG_PORT;
batch_submit_options = getenv("BATCH_OPTIONS");
debug_config(argv[0]);
resources = rmsummary_create(-1);
int c;
while((c = getopt_long(argc, argv, "B:C:F:N:M:T:t:w:W:E:P:S:cd:o:O:vh", long_options, NULL)) > -1) {
switch (c) {
case 'B':
batch_submit_options = xxstrdup(optarg);
break;
case 'C':
config_file = xxstrdup(optarg);
break;
case 'F':
foremen_regex = xxstrdup(optarg);
break;
case 'N':
case 'M':
project_regex = xxstrdup(optarg);
break;
case 'T':
batch_queue_type = batch_queue_type_from_string(optarg);
if(batch_queue_type == BATCH_QUEUE_TYPE_UNKNOWN) {
fprintf(stderr, "unknown batch queue type: %s\n", optarg);
return EXIT_FAILURE;
}
break;
case 't':
worker_timeout = atoi(optarg);
break;
case 'w':
workers_min = atoi(optarg);
break;
case 'W':
workers_max = atoi(optarg);
break;
case LONG_OPT_WORKERS_PER_CYCLE:
workers_per_cycle = atoi(optarg);
break;
case LONG_OPT_TASKS_PER_WORKER:
tasks_per_worker = atof(optarg);
break;
case 'E':
extra_worker_args = xxstrdup(optarg);
break;
case LONG_OPT_CORES:
resources->cores = atoi(optarg);
break;
case LONG_OPT_AMAZON_CREDENTIALS:
amazon_credentials = xxstrdup(optarg);
break;
case LONG_OPT_AMAZON_AMI:
amazon_ami = xxstrdup(optarg);
break;
case LONG_OPT_MEMORY:
resources->memory = atoi(optarg);
break;
case LONG_OPT_DISK:
resources->disk = atoi(optarg);
break;
case LONG_OPT_GPUS:
resources->gpus = atoi(optarg);
break;
case LONG_OPT_AUTOSIZE:
autosize = 1;
break;
case LONG_OPT_FACTORY_TIMEOUT:
factory_timeout = MAX(0, atoi(optarg));
break;
case LONG_OPT_CONDOR_REQUIREMENTS:
if(condor_requirements) {
char *tmp = condor_requirements;
condor_requirements = string_format("(%s && (%s))", tmp, optarg);
free(tmp);
} else {
condor_requirements = string_format("(%s)", optarg);
}
break;
case LONG_OPT_WRAPPER:
wrapper_command = optarg;
break;
case LONG_OPT_WRAPPER_INPUT:
if(!wrapper_input) {
wrapper_input = strdup(optarg);
} else {
wrapper_input = string_format("%s,%s",wrapper_input,optarg);
}
break;
case 'P':
password_file = optarg;
break;
case 'S':
scratch_dir = optarg;
break;
case 'c':
consider_capacity = 1;
break;
case 'd':
debug_flags_set(optarg);
break;
case 'o':
debug_config_file(optarg);
break;
case 'O':
debug_config_file_size(string_metric_parse(optarg));
break;
case 'v':
cctools_version_print(stdout, argv[0]);
exit(EXIT_SUCCESS);
case 'h':
show_help(argv[0]);
exit(EXIT_SUCCESS);
default:
show_help(argv[0]);
return EXIT_FAILURE;
}
}
if(project_regex) {
using_catalog = 1;
}
else if((argc - optind) == 2) {
master_host = argv[optind];
master_port = atoi(argv[optind+1]);
}
else {
fprintf(stderr,"work_queue_factory: You must either give a project name with the -M option or master-name option with a configuration file, or give the master's host and port.\n");
show_help(argv[0]);
exit(1);
}
cctools_version_debug(D_DEBUG, argv[0]);
if(batch_queue_type == BATCH_QUEUE_TYPE_UNKNOWN) {
fprintf(stderr,"work_queue_factory: You must specify a batch type with the -T option.\n");
fprintf(stderr, "valid options:\n");
fprintf(stderr, "%s\n", batch_queue_type_string());
return 1;
}
if(config_file) {
char abs_path_name[PATH_MAX];
if(!realpath(config_file, abs_path_name)) {
fprintf(stderr, "work_queue_factory: could not resolve configuration file path: '%s'.\n", config_file);
exit(EXIT_FAILURE);
}
free(config_file);
/* From now on, read config_file from absolute path */
config_file = xxstrdup(abs_path_name);
if(!read_config_file(config_file)) {
fprintf(stderr,"work_queue_factory: There were errors in the configuration file: %s\n", config_file);
return 1;
}
}
if(workers_min>workers_max) {
fprintf(stderr,"work_queue_factory: min workers (%d) is greater than max workers (%d)\n",workers_min, workers_max);
return 1;
}
/*
Careful here: most of the supported batch systems expect
that jobs are submitting from a single shared filesystem.
Changing to /tmp only works in the case of Condor.
*/
if(!scratch_dir) {
if(batch_queue_type==BATCH_QUEUE_TYPE_CONDOR) {
scratch_dir = string_format("/tmp/wq-pool-%d",getuid());
} else {
scratch_dir = string_format("wq-pool-%d",getuid());
}
}
if(!create_dir(scratch_dir,0777)) {
fprintf(stderr,"work_queue_factory: couldn't create %s: %s",scratch_dir,strerror(errno));
return 1;
}
char cmd[1024];
sprintf(cmd,"cp \"$(which work_queue_worker)\" '%s'",scratch_dir);
if (system(cmd)) {
fprintf(stderr, "work_queue_factory: please add work_queue_worker to your PATH.\n");
exit(EXIT_FAILURE);
}
if(password_file) {
sprintf(cmd,"cp %s %s/pwfile",password_file,scratch_dir);
system(cmd);
}
if(chdir(scratch_dir)!=0) {
fprintf(stderr,"work_queue_factory: couldn't chdir to %s: %s",scratch_dir,strerror(errno));
return 1;
}
signal(SIGINT, handle_abort);
signal(SIGQUIT, handle_abort);
signal(SIGTERM, handle_abort);
signal(SIGHUP, ignore_signal);
queue = batch_queue_create(batch_queue_type);
if(!queue) {
fprintf(stderr,"work_queue_factory: couldn't establish queue type %s",batch_queue_type_to_string(batch_queue_type));
return 1;
}
batch_queue_set_option(queue, "batch-options", batch_submit_options);
batch_queue_set_option(queue, "autosize", autosize ? "yes" : NULL);
set_worker_resources_options( queue );
if (amazon_credentials != NULL) {
batch_queue_set_option(queue, "amazon-credentials", amazon_credentials);
}
if (amazon_ami != NULL) {
batch_queue_set_option(queue, "amazon-ami", amazon_ami);
}
if(condor_requirements != NULL && batch_queue_type != BATCH_QUEUE_TYPE_CONDOR) {
debug(D_NOTICE, "condor_requirements will be ignored as workers will not be running in condor.");
} else {
batch_queue_set_option(queue, "condor-requirements", condor_requirements);
}
mainloop( queue );
batch_queue_delete(queue);
return 0;
}
void category_specify_max_allocation(struct category *c, const struct rmsummary *s) {
rmsummary_delete(c->max_allocation);
c->max_allocation = rmsummary_create(-1);
rmsummary_merge_max(c->max_allocation, s);
}
int category_update_first_allocation(struct category *c, const struct rmsummary *max_worker) {
/* buffer used only for debug output. */
static buffer_t *b = NULL;
if(!b) {
b = malloc(sizeof(buffer_t));
buffer_init(b);
}
if(c->allocation_mode == CATEGORY_ALLOCATION_MODE_FIXED)
return 0;
if(c->total_tasks < 1)
return 0;
struct rmsummary *top = rmsummary_create(-1);
rmsummary_merge_override(top, max_worker);
rmsummary_merge_override(top, c->max_resources_seen);
rmsummary_merge_override(top, c->max_allocation);
if(!c->first_allocation) {
c->first_allocation = rmsummary_create(-1);
}
update_first_allocation_field(c, top, 1, cpu_time);
update_first_allocation_field(c, top, 1, wall_time);
update_first_allocation_field(c, top, c->time_peak_independece, cores);
update_first_allocation_field(c, top, c->time_peak_independece, virtual_memory);
update_first_allocation_field(c, top, c->time_peak_independece, memory);
update_first_allocation_field(c, top, c->time_peak_independece, swap_memory);
update_first_allocation_field(c, top, c->time_peak_independece, bytes_read);
update_first_allocation_field(c, top, c->time_peak_independece, bytes_written);
update_first_allocation_field(c, top, c->time_peak_independece, bytes_received);
update_first_allocation_field(c, top, c->time_peak_independece, bytes_sent);
update_first_allocation_field(c, top, c->time_peak_independece, bandwidth);
update_first_allocation_field(c, top, c->time_peak_independece, total_files);
update_first_allocation_field(c, top, c->time_peak_independece, disk);
update_first_allocation_field(c, top, c->time_peak_independece, max_concurrent_processes);
update_first_allocation_field(c, top, c->time_peak_independece, total_processes);
/* From here on we only print debugging info. */
struct jx *jsum = rmsummary_to_json(c->first_allocation, 1);
if(jsum) {
char *str = jx_print_string(jsum);
debug(D_DEBUG, "Updating first allocation '%s':", c->name);
debug(D_DEBUG, "%s", str);
jx_delete(jsum);
free(str);
}
jsum = rmsummary_to_json(top, 1);
if(jsum) {
char *str = jx_print_string(jsum);
debug(D_DEBUG, "From max resources '%s':", c->name);
debug(D_DEBUG, "%s", str);
jx_delete(jsum);
free(str);
}
rmsummary_delete(top);
return 1;
}
