Rooba
/
boinc
mirror of https://github.com/BOINC/boinc.git
1
0
Fork 0
Code Issues 1 Packages Projects Releases Wiki Activity
boinc/client/work_fetch.cpp

1347 lines
43 KiB
C++
Raw Blame History

// This file is part of BOINC.
// http://boinc.berkeley.edu
// Copyright (C) 2008 University of California
//
// BOINC is free software; you can redistribute it and/or modify it
// under the terms of the GNU Lesser General Public License
// as published by the Free Software Foundation,
// either version 3 of the License, or (at your option) any later version.
//
// BOINC is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
// See the GNU Lesser General Public License for more details.
//
// You should have received a copy of the GNU Lesser General Public License
// along with BOINC. If not, see <http://www.gnu.org/licenses/>.
#include "util.h"
#include "client_state.h"
#include "client_msgs.h"
#include "scheduler_op.h"
#include "work_fetch.h"
using std::vector;
RSC_WORK_FETCH cuda_work_fetch;
RSC_WORK_FETCH ati_work_fetch;
RSC_WORK_FETCH cpu_work_fetch;
WORK_FETCH work_fetch;
#define MIN_BACKOFF_INTERVAL 60
#define MAX_BACKOFF_INTERVAL 86400
// if we ask a project for work for a resource and don't get it,
// we do exponential backoff.
// This constant is an upper bound for this.
// E.g., if we need GPU work, we'll end up asking once a day,
// so if the project develops a GPU app,
// we'll find out about it within a day.
#define FETCH_IF_IDLE_INSTANCE 0
// If resource has an idle instance,
// get work for it from the project with greatest LTD,
// even if it's overworked.
#define FETCH_IF_MAJOR_SHORTFALL 1
// If resource is saturated for less than work_buf_min(),
// get work for it from the project with greatest LTD,
// even if it's overworked.
#define FETCH_IF_MINOR_SHORTFALL 2
// If resource is saturated for less than work_buf_total(),
// get work for it from the non-overworked project with greatest LTD.
#define FETCH_IF_PROJECT_STARVED 3
// If any project is not overworked and has too few jobs
// to use its instance share,
// get work from the one with greatest LTD.
static const char* criterion_name(int criterion) {
switch (criterion) {
case FETCH_IF_IDLE_INSTANCE: return "idle instance";
case FETCH_IF_MAJOR_SHORTFALL: return "major shortfall";
case FETCH_IF_MINOR_SHORTFALL: return "minor shortfall";
case FETCH_IF_PROJECT_STARVED: return "starved";
}
return "unknown";
}
static inline const char* rsc_name(int t) {
switch (t) {
case RSC_TYPE_CPU: return "CPU";
case RSC_TYPE_CUDA: return "NVIDIA GPU";
case RSC_TYPE_ATI: return "ATI GPU";
}
return "Unknown";
}
inline bool prefs_prevent_fetch(PROJECT* p, int rsc_type) {
switch(rsc_type) {
case RSC_TYPE_CPU:
if (p->no_cpu_pref) return true;
break;
case RSC_TYPE_CUDA:
if (p->no_cuda_pref) return true;
break;
case RSC_TYPE_ATI:
if (p->no_ati_pref) return true;
break;
}
return false;
}
// does the project have a downloading or runnable job?
//
static bool has_a_job(PROJECT* p) {
for (unsigned int j=0; j<gstate.results.size(); j++) {
RESULT* rp = gstate.results[j];
if (rp->project != p) continue;
if (rp->state() <= RESULT_FILES_DOWNLOADED) {
return true;
}
}
return false;
}
inline bool has_coproc_app(PROJECT* p, int rsc_type) {
unsigned int i;
for (i=0; i<gstate.app_versions.size(); i++) {
APP_VERSION* avp = gstate.app_versions[i];
if (avp->project != p) continue;
switch(rsc_type) {
case RSC_TYPE_CUDA:
if (avp->ncudas) return true;
break;
case RSC_TYPE_ATI:
if (avp->natis) return true;
break;
}
}
return false;
}
/////////////// RSC_PROJECT_WORK_FETCH ///////////////
bool RSC_PROJECT_WORK_FETCH::compute_may_have_work(PROJECT* p, int rsc_type) {
switch(rsc_type) {
case RSC_TYPE_CPU:
if (p->no_cpu_pref) return false;
break;
case RSC_TYPE_CUDA:
if (p->no_cuda_pref) return false;
if (p->cuda_low_mem) return false;
break;
case RSC_TYPE_ATI:
if (p->no_ati_pref) return false;
if (p->ati_low_mem) return false;
break;
}
return (backoff_time < gstate.now);
}
void RSC_PROJECT_WORK_FETCH::rr_init(PROJECT* p, int rsc_type) {
may_have_work = compute_may_have_work(p, rsc_type);
runnable_share = 0;
fetchable_share = 0;
has_runnable_jobs = false;
sim_nused = 0;
nused_total = 0;
deadlines_missed = 0;
}
// see if the project's debt is beyond what would normally happen;
// if so we conclude that it had a long job that ran in EDF mode;
// avoid asking it for work unless absolutely necessary.
//
bool RSC_PROJECT_WORK_FETCH::overworked() {
double x = gstate.work_buf_total() + gstate.global_prefs.cpu_scheduling_period();
if (x < 86400) x = 86400;
return (long_term_debt < -x);
}
// should this project be accumulating LTD for this resource?
//
bool RSC_PROJECT_WORK_FETCH::debt_eligible(PROJECT* p, RSC_WORK_FETCH& rwf) {
if (p->non_cpu_intensive) return false;
if (p->suspended_via_gui) return false;
if (p->some_result_suspended()) return false;
if (has_runnable_jobs) return true;
// must precede the done_request_more_work check
if (p->dont_request_more_work) return false;
if (backoff_time > gstate.now) return false;
if (prefs_prevent_fetch(p, rwf.rsc_type)) return false;
// NOTE: it's critical that all conditions that might prevent
// us from asking the project for work of this type
// be included in the above list.
// Otherwise we might get in a state where debt accumulates,
// pushing other projects into overworked state
// The last time we asked for work we didn't get any,
// but it's been a while since we asked.
// In this case, accumulate debt until we reach (around) zero, then stop.
//
if (backoff_interval == MAX_BACKOFF_INTERVAL) {
if (long_term_debt > -DEBT_ADJUST_PERIOD) {
return false;
}
}
if (p->min_rpc_time > gstate.now) return false;
return true;
}
void RSC_PROJECT_WORK_FETCH::backoff(PROJECT* p, const char* name) {
if (backoff_interval) {
backoff_interval *= 2;
if (backoff_interval > MAX_BACKOFF_INTERVAL) backoff_interval = MAX_BACKOFF_INTERVAL;
} else {
backoff_interval = MIN_BACKOFF_INTERVAL;
}
double x = drand()*backoff_interval;
backoff_time = gstate.now + x;
if (log_flags.work_fetch_debug) {
msg_printf(p, MSG_INFO,
"[wfd] backing off %s %.0f sec", name, x
);
}
}
/////////////// RSC_WORK_FETCH ///////////////
RSC_PROJECT_WORK_FETCH& RSC_WORK_FETCH::project_state(PROJECT* p) {
switch(rsc_type) {
case RSC_TYPE_CPU: return p->cpu_pwf;
case RSC_TYPE_CUDA: return p->cuda_pwf;
case RSC_TYPE_ATI: return p->ati_pwf;
default: return p->cpu_pwf;
}
}
bool RSC_WORK_FETCH::may_have_work(PROJECT* p) {
if (prefs_prevent_fetch(p, rsc_type)) return false;
RSC_PROJECT_WORK_FETCH& w = project_state(p);
return (w.backoff_time < gstate.now);
}
void RSC_WORK_FETCH::rr_init() {
shortfall = 0;
nidle_now = 0;
sim_nused = 0;
total_fetchable_share = 0;
total_runnable_share = 0;
deadline_missed_instances = 0;
saturated_time = 0;
pending.clear();
busy_time_estimator.reset();
}
void RSC_WORK_FETCH::accumulate_shortfall(double d_time) {
double idle = ninstances - sim_nused;
if (idle > 1e-6) {
shortfall += idle*d_time;
}
#if 0
msg_printf(0, MSG_INFO, "accum shortf (%s): idle %f dt %f sf %f",
rsc_name(rsc_type), idle, d_time, shortfall
);
#endif
}
void RSC_WORK_FETCH::update_saturated_time(double dt) {
double idle = ninstances - sim_nused;
if (idle < 1e-6) {
saturated_time = dt;
}
}
void RSC_WORK_FETCH::update_busy_time(double dur, double nused) {
busy_time_estimator.update(dur, nused);
}
static bool wacky_dcf(PROJECT* p) {
double dcf = p->duration_correction_factor;
return (dcf < 0.02 || dcf > 80.0);
}
// Choose the best project to ask for work for this resource,
// given the specific criterion
//
PROJECT* RSC_WORK_FETCH::choose_project(int criterion) {
PROJECT* pbest = NULL;
switch (criterion) {
case FETCH_IF_IDLE_INSTANCE:
if (nidle_now == 0) return NULL;
break;
case FETCH_IF_MAJOR_SHORTFALL:
if (saturated_time > gstate.work_buf_min()) return NULL;
break;
case FETCH_IF_MINOR_SHORTFALL:
if (saturated_time > gstate.work_buf_total()) return NULL;
break;
case FETCH_IF_PROJECT_STARVED:
if (deadline_missed_instances >= ninstances) return NULL;
break;
}
for (unsigned i=0; i<gstate.projects.size(); i++) {
PROJECT* p = gstate.projects[i];
if (!p->pwf.can_fetch_work) continue;
if (!project_state(p).may_have_work) continue;
RSC_PROJECT_WORK_FETCH& rpwf = project_state(p);
if (rpwf.anon_skip) continue;
switch (criterion) {
case FETCH_IF_MINOR_SHORTFALL:
if (rpwf.overworked()) continue;
if (wacky_dcf(p)) continue;
if (!p->resource_share) continue;
break;
case FETCH_IF_MAJOR_SHORTFALL:
if (wacky_dcf(p)) continue;
if (!p->resource_share) continue;
break;
case FETCH_IF_PROJECT_STARVED:
if (rpwf.overworked()) continue;
if (rpwf.nused_total >= ninstances*rpwf.fetchable_share) continue;
if (!p->resource_share) continue;
break;
}
if (pbest) {
if (!p->resource_share) {
continue;
}
if (pbest->pwf.overall_debt > p->pwf.overall_debt) {
continue;
}
}
pbest = p;
}
if (!pbest) return NULL;
// decide how much work to request from each resource
//
work_fetch.clear_request();
switch (criterion) {
case FETCH_IF_IDLE_INSTANCE:
case FETCH_IF_MAJOR_SHORTFALL:
set_request(pbest, true);
break;
case FETCH_IF_PROJECT_STARVED:
set_request(pbest, false);
break;
case FETCH_IF_MINOR_SHORTFALL:
// in this case, potentially request work for all resources
//
work_fetch.set_all_requests(pbest);
break;
}
// in principle there should be a nonzero request.
// check, just in case
//
if (!req_secs && !req_instances) {
if (log_flags.work_fetch_debug) {
msg_printf(pbest, MSG_INFO,
"[wfd] error: project chosen but zero request"
);
}
return 0;
}
if (log_flags.work_fetch_debug) {
msg_printf(pbest, MSG_INFO,
"chosen: %s %s: %.2f inst, %.2f sec",
criterion_name(criterion), rsc_name(rsc_type),
req_instances, req_secs
);
}
return pbest;
}
// request this project's share of shortfall and instances.
// don't request anything if project is overworked or backed off.
//
void RSC_WORK_FETCH::set_request(PROJECT* p, bool allow_overworked) {
RSC_PROJECT_WORK_FETCH& w = project_state(p);
if (!w.may_have_work) return;
if (w.anon_skip) return;
if (!allow_overworked && w.overworked()) return;
if (shortfall) {
if (wacky_dcf(p)) {
// if project's DCF is too big or small,
// its completion time estimates are useless; just ask for 1 second
//
req_secs = 1;
} else {
req_secs = shortfall * w.fetchable_share;
}
}
// the number of additional instances needed to have our share
//
double x1 = (ninstances * w.fetchable_share) - w.nused_total;
// our share of the idle instances
//
double x2 = nidle_now * w.fetchable_share;
req_instances = std::max(x1, x2);
if (req_instances && !req_secs) {
req_secs = 1;
}
}
void RSC_WORK_FETCH::print_state(const char* name) {
msg_printf(0, MSG_INFO,
"[wfd] %s: shortfall %.2f nidle %.2f saturated %.2f busy %.2f RS fetchable %.2f runnable %.2f",
name,
shortfall, nidle_now, saturated_time, busy_time_estimator.get_busy_time(),
total_fetchable_share, total_runnable_share
);
for (unsigned int i=0; i<gstate.projects.size(); i++) {
PROJECT* p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
RSC_PROJECT_WORK_FETCH& pwf = project_state(p);
double bt = pwf.backoff_time>gstate.now?pwf.backoff_time-gstate.now:0;
bool blocked_by_prefs = false;
switch (rsc_type) {
case RSC_TYPE_CPU:
if (p->no_cpu_pref) blocked_by_prefs = true;
break;
case RSC_TYPE_CUDA:
if (p->no_cuda_pref) blocked_by_prefs = true;
break;
case RSC_TYPE_ATI:
if (p->no_ati_pref) blocked_by_prefs = true;
break;
}
msg_printf(p, MSG_INFO,
"[wfd] %s: fetch share %.2f LTD %.2f backoff dt %.2f int %.2f%s%s%s%s%s%s%s",
name,
pwf.fetchable_share, pwf.long_term_debt, bt, pwf.backoff_interval,
p->suspended_via_gui?" (susp via GUI)":"",
p->master_url_fetch_pending?" (master fetch pending)":"",
p->min_rpc_time > gstate.now?" (comm deferred)":"",
p->dont_request_more_work?" (no new tasks)":"",
pwf.overworked()?" (overworked)":"",
p->too_many_uploading_results?" (too many uploads)":"",
blocked_by_prefs?" (blocked by prefs)":""
);
}
}
void RSC_WORK_FETCH::clear_request() {
req_secs = 0;
req_instances = 0;
}
// update long-term debts for a resource.
//
void RSC_WORK_FETCH::update_long_term_debts() {
unsigned int i;
int neligible = 0;
double ders = 0;
PROJECT* p;
// find the total resource share of eligible projects
//
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
RSC_PROJECT_WORK_FETCH& w = project_state(p);
if (w.debt_eligible(p, *this)) {
ders += p->resource_share;
neligible++;
}
}
if (!neligible) {
if (log_flags.debt_debug) {
msg_printf(0, MSG_INFO,
"[debt] %s: no eligible projects", rsc_name(rsc_type)
);
}
return;
}
double max_debt=0;
bool first = true;
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
if (!p->resource_share) continue;
RSC_PROJECT_WORK_FETCH& w = project_state(p);
if (w.debt_eligible(p, *this)) {
double share_frac = p->resource_share/ders;
// the change to a project's debt is:
// (how much it's owed) - (how much it got)
//
double delta = share_frac*secs_this_debt_interval - w.secs_this_debt_interval;
delta /= ninstances;
w.long_term_debt += delta;
if (log_flags.debt_debug) {
msg_printf(p, MSG_INFO,
"[debt] %s LTD %.2f delta %.2f (%.2f*%.2f - %.2f)/%d",
rsc_name(rsc_type),
w.long_term_debt, delta, share_frac,
secs_this_debt_interval,
w.secs_this_debt_interval,
ninstances
);
}
if (first) {
max_debt = w.long_term_debt;
first = false;
} else {
if (w.long_term_debt > max_debt) {
max_debt = w.long_term_debt;
}
}
} else {
if (log_flags.debt_debug) {
msg_printf(p, MSG_INFO,
"[debt] %s ineligible; LTD %.2f",
rsc_name(rsc_type), w.long_term_debt
);
}
}
}
// The net change may be
// - positive if the resource wasn't fully utilized during the debt interval
// - negative it was overcommitted (e.g., CPU)
// We need to keep eligible projects from diverging from non-eligible ones;
// also, if all the debts are large negative we need to gradually
// shift them towards zero.
// To do this, we add an offset as follows:
// delta_limit is the largest rate at which any project's debt
// could increase or decrease.
// If the largest debt is close to zero (relative to delta_limit)
// than add an offset that will bring it exactly to zero.
// Otherwise add an offset of 2*delta_limit,
// which will gradually bring all the debts towards zero
//
// The policy of keeping the max debt at zero is important;
// it means that new projects will begin in parity with high-debt project,
// and won't wait for months to get work.
//
double offset;
double delta_limit = secs_this_debt_interval;
if (max_debt > -2*delta_limit) {
if (fabs(max_debt) < 1e-6) max_debt = 0;
offset = max_debt?-max_debt:0; // avoid -0
} else {
offset = 2*delta_limit;
}
if (log_flags.debt_debug) {
msg_printf(0, MSG_INFO, "[debt] %s LTD: adding offset %f",
rsc_name(rsc_type), offset
);
}
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
if (!p->resource_share) continue;
RSC_PROJECT_WORK_FETCH& w = project_state(p);
if (w.debt_eligible(p, *this)) {
w.long_term_debt += offset;
} else {
if (offset > 0) {
w.long_term_debt += offset;
}
}
if (w.long_term_debt > 0) w.long_term_debt = 0;
}
}
// update short-term debts for a resource.
//
void RSC_WORK_FETCH::update_short_term_debts() {
unsigned int i;
PROJECT* p;
int nprojects=0, nrprojects=0;
double share_frac;
double total_short_term_debt = 0;
double rrs = gstate.runnable_resource_share(rsc_type);
for (i=0; i<gstate.projects.size(); i++) {
double delta;
p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
if (!p->resource_share) continue;
RSC_PROJECT_WORK_FETCH& rpwf = project_state(p);
nprojects++;
if (p->runnable(rsc_type)) {
nrprojects++;
share_frac = p->resource_share/rrs;
delta = share_frac*secs_this_debt_interval
- rpwf.secs_this_debt_interval;
delta /= ninstances;
if (log_flags.std_debug) {
msg_printf(p, MSG_INFO,
"[std_debug] %s STD delta %.2f (%.2f*%.2f - %.2f)/%d",
rsc_name(rsc_type),
delta,
share_frac,
secs_this_debt_interval,
rpwf.secs_this_debt_interval,
ninstances
);
}
rpwf.short_term_debt += delta;
}
total_short_term_debt += rpwf.short_term_debt;
}
// normalize so mean is zero, and limit abs value to MAX_STD
//
if (nrprojects) {
double avg_short_term_debt = total_short_term_debt / nprojects;
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
if (!p->resource_share) continue;
RSC_PROJECT_WORK_FETCH& rpwf = project_state(p);
rpwf.short_term_debt -= avg_short_term_debt;
if (rpwf.short_term_debt > MAX_STD) {
rpwf.short_term_debt = MAX_STD;
}
if (rpwf.short_term_debt < -MAX_STD) {
rpwf.short_term_debt = -MAX_STD;
}
if (p->runnable(rsc_type)) {
if (log_flags.std_debug) {
msg_printf(p, MSG_INFO,
"[std_debug] %s STD %.2f",
rsc_name(rsc_type), rpwf.short_term_debt
);
}
}
}
}
}
/////////////// PROJECT_WORK_FETCH ///////////////
bool PROJECT_WORK_FETCH::compute_can_fetch_work(PROJECT* p) {
if (p->non_cpu_intensive) return false;
if (p->suspended_via_gui) return false;
if (p->master_url_fetch_pending) return false;
if (p->min_rpc_time > gstate.now) return false;
if (p->dont_request_more_work) return false;
if (p->some_download_stalled()) return false;
if (p->some_result_suspended()) return false;
if (p->too_many_uploading_results) return false;
return true;
}
void PROJECT_WORK_FETCH::reset(PROJECT* p) {
p->cpu_pwf.reset();
p->cuda_pwf.reset();
p->ati_pwf.reset();
}
/////////////// WORK_FETCH ///////////////
void WORK_FETCH::rr_init() {
cpu_work_fetch.rr_init();
// do these even if no device; there may be "coproc_missing" jobs
cuda_work_fetch.rr_init();
ati_work_fetch.rr_init();
for (unsigned int i=0; i<gstate.projects.size(); i++) {
PROJECT* p = gstate.projects[i];
p->pwf.can_fetch_work = p->pwf.compute_can_fetch_work(p);
p->pwf.has_runnable_jobs = false;
p->cpu_pwf.rr_init(p, RSC_TYPE_CPU);
if (coproc_cuda) {
p->cuda_pwf.rr_init(p, RSC_TYPE_CUDA);
}
if (coproc_ati) {
p->ati_pwf.rr_init(p, RSC_TYPE_ATI);
}
}
}
void WORK_FETCH::set_all_requests(PROJECT* p) {
cpu_work_fetch.set_request(p, false);
if (coproc_cuda && gpus_usable) {
cuda_work_fetch.set_request(p, false);
}
if (coproc_ati && gpus_usable) {
ati_work_fetch.set_request(p, false);
}
}
// Compute an "overall long-term debt" for each project.
// This is a sum of per-resource terms, scaled by the relative speed of the resource.
// The term for a resource is its LTD plus an estimate of queued work.
//
void WORK_FETCH::set_overall_debts() {
unsigned int i;
PROJECT* p;
RESULT* rp;
APP_VERSION* avp;
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
p->cpu_pwf.queue_est = 0;
p->cuda_pwf.queue_est = 0;
p->ati_pwf.queue_est = 0;
}
for (i=0; i<gstate.results.size(); i++) {
rp = gstate.results[i];
p = rp->project;
if (!rp->nearly_runnable()) continue;
if (p->non_cpu_intensive) continue;
double dt = rp->estimated_time_remaining(false);
avp = rp->avp;
p->cpu_pwf.queue_est += dt*avp->avg_ncpus;
p->cuda_pwf.queue_est += dt*avp->ncudas;
p->ati_pwf.queue_est += dt*avp->natis;
}
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
double queue_debt = p->cpu_pwf.queue_est/gstate.ncpus;
p->pwf.overall_debt = p->cpu_pwf.long_term_debt - queue_debt;
if (coproc_cuda) {
p->pwf.overall_debt += cuda_work_fetch.relative_speed*
(p->cuda_pwf.long_term_debt - p->cuda_pwf.queue_est/coproc_cuda->count);
}
if (coproc_ati) {
p->pwf.overall_debt += ati_work_fetch.relative_speed*
(p->ati_pwf.long_term_debt - p->ati_pwf.queue_est/coproc_ati->count);
}
}
}
void WORK_FETCH::zero_debts() {
for (unsigned i=0; i<gstate.projects.size(); i++) {
PROJECT* p = gstate.projects[i];
p->cpu_pwf.zero_debt();
p->cuda_pwf.zero_debt();
p->ati_pwf.zero_debt();
}
}
void WORK_FETCH::print_state() {
msg_printf(0, MSG_INFO, "[wfd] ------- start work fetch state -------");
msg_printf(0, MSG_INFO, "[wfd] target work buffer: %.2f + %.2f sec",
gstate.work_buf_min(), gstate.work_buf_additional()
);
cpu_work_fetch.print_state("CPU");
if (coproc_cuda) {
cuda_work_fetch.print_state("NVIDIA GPU");
}
if (coproc_ati) {
ati_work_fetch.print_state("ATI GPU");
}
for (unsigned int i=0; i<gstate.projects.size(); i++) {
PROJECT* p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
msg_printf(p, MSG_INFO, "[wfd] overall LTD %.2f", p->pwf.overall_debt);
}
msg_printf(0, MSG_INFO, "[wfd] ------- end work fetch state -------");
}
void WORK_FETCH::clear_request() {
cpu_work_fetch.clear_request();
cuda_work_fetch.clear_request();
ati_work_fetch.clear_request();
}
// we're going to contact this project for reasons other than work fetch;
// decide if we should piggy-back a work fetch request.
//
void WORK_FETCH::compute_work_request(PROJECT* p) {
clear_request();
if (p->dont_request_more_work) return;
if (p->non_cpu_intensive) {
if (!has_a_job(p)) {
cpu_work_fetch.req_secs = 1;
}
return;
}
// See if this is the project we'd ask for work anyway.
// Temporarily clear resource backoffs,
// since we're going to contact this project in any case.
//
double cpu_save = p->cpu_pwf.backoff_time;
double cuda_save = p->cuda_pwf.backoff_time;
double ati_save = p->ati_pwf.backoff_time;
p->cpu_pwf.backoff_time = 0;
p->cuda_pwf.backoff_time = 0;
p->ati_pwf.backoff_time = 0;
PROJECT* pbest = choose_project();
p->cpu_pwf.backoff_time = cpu_save;
p->cuda_pwf.backoff_time = cuda_save;
p->ati_pwf.backoff_time = ati_save;
if (p == pbest) {
// Ask for work for all devices w/ a shortfall.
// Otherwise we can have a situation where a GPU is idle,
// we ask only for GPU work, and the project never has any
//
work_fetch.set_all_requests(pbest);
return;
}
// if not, don't request any work
//
clear_request();
}
// see if there's a fetchable non-CPU-intensive project without work
//
PROJECT* WORK_FETCH::non_cpu_intensive_project_needing_work() {
for (unsigned int i=0; i<gstate.projects.size(); i++) {
PROJECT* p = gstate.projects[i];
if (!p->non_cpu_intensive) continue;
if (!p->can_request_work()) continue;
if (p->cpu_pwf.backoff_time > gstate.now) continue;
if (has_a_job(p)) continue;
clear_request();
cpu_work_fetch.req_secs = 1;
return p;
}
return 0;
}
// choose a project to fetch work from,
// and set the request fields of resource objects
//
PROJECT* WORK_FETCH::choose_project() {
PROJECT* p = 0;
if (log_flags.work_fetch_debug) {
msg_printf(0, MSG_INFO, "[wfd]: work fetch start");
}
p = non_cpu_intensive_project_needing_work();
if (p) return p;
gstate.compute_nuploading_results();
gstate.rr_simulation();
set_overall_debts();
bool cuda_usable = coproc_cuda && gpus_usable;
bool ati_usable = coproc_ati && gpus_usable;
if (cuda_usable) {
p = cuda_work_fetch.choose_project(FETCH_IF_IDLE_INSTANCE);
}
if (!p && ati_usable) {
p = ati_work_fetch.choose_project(FETCH_IF_IDLE_INSTANCE);
}
if (!p) {
p = cpu_work_fetch.choose_project(FETCH_IF_IDLE_INSTANCE);
}
if (!p && cuda_usable) {
p = cuda_work_fetch.choose_project(FETCH_IF_MAJOR_SHORTFALL);
}
if (!p && ati_usable) {
p = ati_work_fetch.choose_project(FETCH_IF_MAJOR_SHORTFALL);
}
if (!p) {
p = cpu_work_fetch.choose_project(FETCH_IF_MAJOR_SHORTFALL);
}
if (!p && cuda_usable) {
p = cuda_work_fetch.choose_project(FETCH_IF_MINOR_SHORTFALL);
}
if (!p && ati_usable) {
p = ati_work_fetch.choose_project(FETCH_IF_MINOR_SHORTFALL);
}
if (!p) {
p = cpu_work_fetch.choose_project(FETCH_IF_MINOR_SHORTFALL);
}
#if 0
// don't try to maintain GPU work for all projects,
// since we don't use round-robin scheduling for GPUs
//
if (!p && cuda_usable) {
p = cuda_work_fetch.choose_project(FETCH_IF_PROJECT_STARVED);
}
if (!p && ati_usable) {
p = ati_work_fetch.choose_project(FETCH_IF_PROJECT_STARVED);
}
#endif
if (!p) {
p = cpu_work_fetch.choose_project(FETCH_IF_PROJECT_STARVED);
}
if (log_flags.work_fetch_debug) {
print_state();
if (!p) {
msg_printf(0, MSG_INFO, "[wfd] No project chosen for work fetch");
}
}
return p;
}
void WORK_FETCH::accumulate_inst_sec(ACTIVE_TASK* atp, double dt) {
APP_VERSION* avp = atp->result->avp;
PROJECT* p = atp->result->project;
double x = dt*avp->avg_ncpus;
p->cpu_pwf.secs_this_debt_interval += x;
cpu_work_fetch.secs_this_debt_interval += x;
if (coproc_cuda) {
x = dt*avp->ncudas;
p->cuda_pwf.secs_this_debt_interval += x;
cuda_work_fetch.secs_this_debt_interval += x;
}
if (coproc_ati) {
x = dt*avp->natis;
p->ati_pwf.secs_this_debt_interval += x;
ati_work_fetch.secs_this_debt_interval += x;
}
}
// find total and per-project resource shares for each resource
//
void WORK_FETCH::compute_shares() {
unsigned int i;
PROJECT* p;
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
if (p->cpu_pwf.has_runnable_jobs) {
cpu_work_fetch.total_runnable_share += p->resource_share;
}
if (p->cuda_pwf.has_runnable_jobs) {
cuda_work_fetch.total_runnable_share += p->resource_share;
}
if (p->ati_pwf.has_runnable_jobs) {
ati_work_fetch.total_runnable_share += p->resource_share;
}
if (!p->pwf.can_fetch_work) continue;
if (p->cpu_pwf.may_have_work) {
cpu_work_fetch.total_fetchable_share += p->resource_share;
}
if (coproc_cuda && p->cuda_pwf.may_have_work) {
cuda_work_fetch.total_fetchable_share += p->resource_share;
}
if (coproc_ati && p->ati_pwf.may_have_work) {
ati_work_fetch.total_fetchable_share += p->resource_share;
}
}
for (i=0; i<gstate.projects.size(); i++) {
p = gstate.projects[i];
if (p->non_cpu_intensive) continue;
if (p->cpu_pwf.has_runnable_jobs) {
p->cpu_pwf.runnable_share = p->resource_share/cpu_work_fetch.total_runnable_share;
}
if (p->cuda_pwf.has_runnable_jobs) {
p->cuda_pwf.runnable_share = p->resource_share/cuda_work_fetch.total_runnable_share;
}
if (p->ati_pwf.has_runnable_jobs) {
p->ati_pwf.runnable_share = p->resource_share/ati_work_fetch.total_runnable_share;
}
if (!p->pwf.can_fetch_work) continue;
if (p->cpu_pwf.may_have_work) {
p->cpu_pwf.fetchable_share = cpu_work_fetch.total_fetchable_share?p->resource_share/cpu_work_fetch.total_fetchable_share:1;
if (log_flags.work_fetch_debug) {
msg_printf(p, MSG_INFO,
"[wfd] FS: %f = %f/%f\n",
p->cpu_pwf.fetchable_share, p->resource_share,
cpu_work_fetch.total_fetchable_share
);
}
}
if (coproc_cuda && p->cuda_pwf.may_have_work) {
p->cuda_pwf.fetchable_share = cuda_work_fetch.total_fetchable_share?p->resource_share/cuda_work_fetch.total_fetchable_share:1;
}
if (coproc_ati && p->ati_pwf.may_have_work) {
p->ati_pwf.fetchable_share = ati_work_fetch.total_fetchable_share?p->resource_share/ati_work_fetch.total_fetchable_share:1;
}
}
}
void WORK_FETCH::write_request(FILE* f, PROJECT* p) {
double work_req = cpu_work_fetch.req_secs;
// if project is anonymous platform, set the overall work req
// to the max of the requests of resource types for which we have versions.
// Otherwise projects with old schedulers won't send us work.
// THIS CAN BE REMOVED AT SOME POINT
//
if (p->anonymous_platform) {
if (has_coproc_app(p, RSC_TYPE_CUDA)) {
if (cuda_work_fetch.req_secs > work_req) {
work_req = cuda_work_fetch.req_secs;
}
}
if (has_coproc_app(p, RSC_TYPE_ATI)) {
if (ati_work_fetch.req_secs > work_req) {
work_req = ati_work_fetch.req_secs;
}
}
}
fprintf(f,
" <work_req_seconds>%f</work_req_seconds>\n"
" <cpu_req_secs>%f</cpu_req_secs>\n"
" <cpu_req_instances>%f</cpu_req_instances>\n"
" <estimated_delay>%f</estimated_delay>\n",
work_req,
cpu_work_fetch.req_secs,
cpu_work_fetch.req_instances,
cpu_work_fetch.req_secs?cpu_work_fetch.busy_time_estimator.get_busy_time():0
);
if (log_flags.work_fetch_debug) {
char buf[256], buf2[256];
sprintf(buf,
"[wfd] request: %.2f sec CPU (%.2f sec, %.2f)",
work_req,
cpu_work_fetch.req_secs, cpu_work_fetch.req_instances
);
if (coproc_cuda) {
sprintf(buf2, " NVIDIA GPU (%.2f sec, %.2f)",
cuda_work_fetch.req_secs, cuda_work_fetch.req_instances
);
strcat(buf, buf2);
}
if (coproc_ati) {
sprintf(buf2, " ATI GPU (%.2f sec, %.2f)",
ati_work_fetch.req_secs, ati_work_fetch.req_instances
);
strcat(buf, buf2);
}
msg_printf(p, MSG_INFO, buf);
}
}
// we just got a scheduler reply with the given jobs; update backoffs
//
void WORK_FETCH::handle_reply(
PROJECT* p, SCHEDULER_REPLY* srp, vector<RESULT*> new_results
) {
unsigned int i;
bool got_cpu = false, got_cuda = false, got_ati = false;
// handle project-supplied backoff requests
//
if (srp->cpu_backoff) {
p->cpu_pwf.backoff_time = gstate.now + srp->cpu_backoff;
}
if (srp->cuda_backoff) {
p->cuda_pwf.backoff_time = gstate.now + srp->cuda_backoff;
}
if (srp->ati_backoff) {
p->ati_pwf.backoff_time = gstate.now + srp->ati_backoff;
}
// if didn't get any jobs, back off on requested resource types
//
if (!new_results.size()) {
// but not if RPC was requested by project
//
if (p->sched_rpc_pending != RPC_REASON_PROJECT_REQ) {
if (cpu_work_fetch.req_secs && !srp->cpu_backoff) {
p->cpu_pwf.backoff(p, "CPU");
}
if (coproc_cuda && coproc_cuda->req_secs && !srp->cuda_backoff) {
p->cuda_pwf.backoff(p, "NVIDIA GPU");
}
if (coproc_ati && coproc_ati->req_secs && !srp->ati_backoff) {
p->ati_pwf.backoff(p, "ATI GPU");
}
}
return;
}
// if we did get jobs, clear backoff on resource types
//
for (i=0; i<new_results.size(); i++) {
RESULT* rp = new_results[i];
if (rp->avp->ncudas) got_cuda = true;
else if (rp->avp->natis) got_ati = true;
else got_cpu = true;
}
if (got_cpu) p->cpu_pwf.clear_backoff();
if (got_cuda) p->cuda_pwf.clear_backoff();
if (got_ati) p->ati_pwf.clear_backoff();
}
// set up for initial RPC.
// arrange to always get one job, even if we don't need it or can't handle it.
// (this is probably what user wants)
//
void WORK_FETCH::set_initial_work_request() {
cpu_work_fetch.req_secs = 1;
cpu_work_fetch.req_instances = 0;
cpu_work_fetch.busy_time_estimator.reset();
if (coproc_cuda) {
cuda_work_fetch.req_secs = 1;
cuda_work_fetch.req_instances = 0;
cuda_work_fetch.busy_time_estimator.reset();
}
if (coproc_ati) {
ati_work_fetch.req_secs = 1;
ati_work_fetch.req_instances = 0;
ati_work_fetch.busy_time_estimator.reset();
}
}
// called once, at client startup
//
void WORK_FETCH::init() {
cpu_work_fetch.init(RSC_TYPE_CPU, gstate.ncpus, 1);
double cpu_flops = gstate.ncpus*gstate.host_info.p_fpops;
// use 20% as a rough estimate of GPU efficiency
if (coproc_cuda) {
cuda_work_fetch.init(
RSC_TYPE_CUDA, coproc_cuda->count,
coproc_cuda->count*0.2*coproc_cuda->peak_flops()/cpu_flops
);
}
if (coproc_ati) {
ati_work_fetch.init(
RSC_TYPE_ATI,
coproc_ati->count,
coproc_ati->count*0.2*coproc_ati->peak_flops()/cpu_flops
);
}
if (config.zero_debts) {
zero_debts();
}
// see what resources anon platform projects can use
//
unsigned int i, j;
for (i=0; i<gstate.projects.size(); i++) {
PROJECT* p = gstate.projects[i];
if (!p->anonymous_platform) continue;
p->cpu_pwf.anon_skip = true;
p->cuda_pwf.anon_skip = true;
p->ati_pwf.anon_skip = true;
for (j=0; j<gstate.app_versions.size(); j++) {
APP_VERSION* avp = gstate.app_versions[j];
if (avp->project != p) continue;
if (avp->ncudas) {
p->cuda_pwf.anon_skip = false;
} else if (avp->natis) {
p->ati_pwf.anon_skip = false;
} else {
p->cpu_pwf.anon_skip = false;
}
}
}
}
////////////////////////
void CLIENT_STATE::compute_nuploading_results() {
unsigned int i;
for (i=0; i<projects.size(); i++) {
projects[i]->nuploading_results = 0;
projects[i]->too_many_uploading_results = false;
}
for (i=0; i<results.size(); i++) {
RESULT* rp = results[i];
if (rp->state() == RESULT_FILES_UPLOADING) {
rp->project->nuploading_results++;
}
}
int n = gstate.ncpus;
if (coproc_cuda && coproc_cuda->count > n) {
n = coproc_cuda->count;
}
if (coproc_ati && coproc_ati->count > n) {
n = coproc_ati->count;
}
n *= 2;
for (i=0; i<projects.size(); i++) {
if (projects[i]->nuploading_results > n) {
projects[i]->too_many_uploading_results = true;
}
}
}
bool PROJECT::runnable(int rsc_type) {
if (suspended_via_gui) return false;
for (unsigned int i=0; i<gstate.results.size(); i++) {
RESULT* rp = gstate.results[i];
if (rp->project != this) continue;
switch (rsc_type) {
case RSC_TYPE_ANY:
break;
case RSC_TYPE_CPU:
if (rp->uses_coprocs()) continue;
break;
case RSC_TYPE_CUDA:
if (rp->avp->ncudas == 0) continue;
break;
case RSC_TYPE_ATI:
if (rp->avp->natis == 0) continue;
break;
}
if (rp->runnable()) return true;
}
return false;
}
bool PROJECT::downloading() {
if (suspended_via_gui) return false;
for (unsigned int i=0; i<gstate.results.size(); i++) {
RESULT* rp = gstate.results[i];
if (rp->project != this) continue;
if (rp->downloading()) return true;
}
return false;
}
bool PROJECT::some_result_suspended() {
unsigned int i;
for (i=0; i<gstate.results.size(); i++) {
RESULT *rp = gstate.results[i];
if (rp->project != this) continue;
if (rp->suspended_via_gui) return true;
}
return false;
}
bool PROJECT::can_request_work() {
if (suspended_via_gui) return false;
if (master_url_fetch_pending) return false;
if (min_rpc_time > gstate.now) return false;
if (dont_request_more_work) return false;
if (gstate.in_abort_sequence) return false;
return true;
}
bool PROJECT::potentially_runnable() {
if (runnable(RSC_TYPE_ANY)) return true;
if (can_request_work()) return true;
if (downloading()) return true;
return false;
}
bool PROJECT::nearly_runnable() {
if (runnable(RSC_TYPE_ANY)) return true;
if (downloading()) return true;
return false;
}
bool RESULT::runnable() {
if (suspended_via_gui) return false;
if (project->suspended_via_gui) return false;
if (state() != RESULT_FILES_DOWNLOADED) return false;
if (coproc_missing) return false;
if (schedule_backoff > gstate.now) return false;
return true;
}
bool RESULT::nearly_runnable() {
return runnable() || downloading();
}
// Return true if the result is waiting for its files to download,
// and nothing prevents this from happening soon
//
bool RESULT::downloading() {
if (suspended_via_gui) return false;
if (project->suspended_via_gui) return false;
if (state() > RESULT_FILES_DOWNLOADING) return false;
return true;
}
double RESULT::estimated_duration_uncorrected() {
return wup->rsc_fpops_est/avp->flops;
}
// estimate how long a result will take on this host
//
#ifdef SIM
double RESULT::estimated_duration(bool for_work_fetch) {
SIM_PROJECT* spp = (SIM_PROJECT*)project;
if (dual_dcf && for_work_fetch && spp->completions_ratio_mean) {
return estimated_duration_uncorrected()*spp->completions_ratio_mean;
}
return estimated_duration_uncorrected()*project->duration_correction_factor;
}
#else
double RESULT::estimated_duration(bool) {
return estimated_duration_uncorrected()*project->duration_correction_factor;
}
#endif
double RESULT::estimated_time_remaining(bool for_work_fetch) {
if (computing_done()) return 0;
ACTIVE_TASK* atp = gstate.lookup_active_task_by_result(this);
if (atp) {
return atp->est_dur(for_work_fetch) - atp->elapsed_time;
}
return estimated_duration(for_work_fetch);
}
// Returns the estimated total elapsed time of this task.
// Compute this as a weighted average of estimates based on
// 1) the workunit's flops count (static estimate)
// 2) the current elapsed time and fraction done (dynamic estimate)
//
double ACTIVE_TASK::est_dur(bool for_work_fetch) {
if (fraction_done >= 1) return 0;
double wu_est = result->estimated_duration(for_work_fetch);
if (fraction_done <= 0) return wu_est;
if (wu_est < elapsed_time) wu_est = elapsed_time;
double frac_est = elapsed_time / fraction_done;
double fraction_left = 1-fraction_done;
double wu_weight = fraction_left * fraction_left * fraction_left;
double fd_weight = 1 - wu_weight;
double x = fd_weight*frac_est + wu_weight*wu_est;
#if 1
if (log_flags.rr_simulation) {
msg_printf(result->project, MSG_INFO,
"[rr_sim] %s dur: %.2f = %.3f*%.2f + %.3f*%.2f",
result->name, x, fd_weight, frac_est, wu_weight, wu_est
);
}
#endif
return x;
}
// the fraction of time BOINC is processing
//
double CLIENT_STATE::overall_cpu_frac() {
double running_frac = time_stats.on_frac * time_stats.active_frac;
if (running_frac < 0.01) running_frac = 0.01;
if (running_frac > 1) running_frac = 1;
return running_frac;
}
// called when benchmarks change
//
void CLIENT_STATE::scale_duration_correction_factors(double factor) {
if (factor <= 0) return;
for (unsigned int i=0; i<projects.size(); i++) {
PROJECT* p = projects[i];
p->duration_correction_factor *= factor;
}
if (log_flags.dcf_debug) {
msg_printf(NULL, MSG_INFO,
"[dcf] scaling all duration correction factors by %f",
factor
);
}
}
// Choose a new host CPID.
// If using account manager, do scheduler RPCs
// to all acct-mgr-attached projects to propagate the CPID
//
void CLIENT_STATE::generate_new_host_cpid() {
host_info.generate_host_cpid();
for (unsigned int i=0; i<projects.size(); i++) {
if (projects[i]->attached_via_acct_mgr) {
projects[i]->sched_rpc_pending = RPC_REASON_ACCT_MGR_REQ;
projects[i]->set_min_rpc_time(now + 15, "Sending new host CPID");
}
}
}
Reference in New Issue View Git Blame Copy Permalink