boinc/sched/credit.cpp

1126 lines
36 KiB
C++
Raw Normal View History

2009-08-11 15:17:37 +00:00
// 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/>.
// Modify claimed credit based on the historical granted credit if
// the project is configured to do this
//
// functions related to the computation and granting of credit
// Note: this is credit.cpp rather than sched_credit.cpp
// because you might grant credit e.g. from a trickle handler
#include <cmath>
2009-08-11 15:17:37 +00:00
#include "boinc_db.h"
#include "error_numbers.h"
2009-08-11 15:17:37 +00:00
#include "sched_config.h"
#include "sched_customize.h"
2009-08-11 15:17:37 +00:00
#include "sched_msgs.h"
#include "sched_util.h"
#include "sched_shmem.h"
#include "sched_types.h"
2009-08-11 15:17:37 +00:00
#include "credit.h"
double fpops_to_credit(double fpops) {
return fpops*COBBLESTONE_SCALE;
}
double cpu_time_to_credit(double cpu_time, double cpu_flops_sec) {
return fpops_to_credit(cpu_time*cpu_flops_sec);
2009-08-11 15:17:37 +00:00
}
// Grant the host (and associated user and team)
// the given amount of credit for work that started at the given time.
// Update the user and team records,
// but not the host record (caller must update)
//
int grant_credit(DB_HOST &host, double start_time, double credit) {
DB_USER user;
DB_TEAM team;
int retval;
char buf[256];
double now = dtime();
// first, process the host
update_average(
now,
start_time, credit, CREDIT_HALF_LIFE,
host.expavg_credit, host.expavg_time
);
host.total_credit += credit;
// then the user
retval = user.lookup_id(host.userid);
if (retval) {
log_messages.printf(MSG_CRITICAL,
"lookup of user %lu failed: %s\n",
host.userid, boincerror(retval)
);
return retval;
}
update_average(
now,
start_time, credit, CREDIT_HALF_LIFE,
user.expavg_credit, user.expavg_time
);
sprintf(
buf, "total_credit=total_credit+%.15e, expavg_credit=%.15e, expavg_time=%.15e",
credit, user.expavg_credit, user.expavg_time
);
retval = user.update_field(buf);
if (retval) {
log_messages.printf(MSG_CRITICAL,
"update of user %lu failed: %s\n",
host.userid, boincerror(retval)
);
}
// and finally the team
if (user.teamid) {
retval = team.lookup_id(user.teamid);
if (retval) {
log_messages.printf(MSG_CRITICAL,
"lookup of team %lu failed: %s\n",
user.teamid, boincerror(retval)
);
return retval;
}
update_average(
now,
start_time, credit, CREDIT_HALF_LIFE,
team.expavg_credit, team.expavg_time
);
sprintf(buf,
"total_credit=total_credit+%.15e, expavg_credit=%.15e, expavg_time=%.15e",
credit, team.expavg_credit, team.expavg_time
);
retval = team.update_field(buf);
if (retval) {
log_messages.printf(MSG_CRITICAL,
"update of team %lu failed: %s\n",
team.id, boincerror(retval)
);
}
}
return 0;
}
int grant_credit_by_app(RESULT& result, double credit) {
DB_CREDIT_USER cu;
char clause1[1024], clause2[1024];
double now = dtime();
sprintf(clause1, "where userid=%lu and appid=%lu", result.userid, result.appid);
int retval = cu.lookup(clause1);
if (retval) {
cu.clear();
cu.userid = result.userid;
cu.appid = result.appid;
cu.expavg_time = dtime();
retval = cu.insert();
if (retval) return retval;
}
update_average(
now,
result.sent_time,
credit, CREDIT_HALF_LIFE,
cu.expavg, cu.expavg_time
);
sprintf(clause1,
"total=total+%.15e, expavg=%.15e, expavg_time=%.15e, njobs=njobs+1",
credit, cu.expavg, cu.expavg_time
);
sprintf(clause2,
"userid=%lu and appid=%lu", result.userid, result.appid
);
retval = cu.update_fields_noid(clause1, clause2);
if (retval) return retval;
// team. keep track of credit for zero (no team) also
//
DB_CREDIT_TEAM ct;
sprintf(clause1, "where teamid=%lu and appid=%lu", result.teamid, result.appid);
retval = ct.lookup(clause1);
if (retval) {
ct.clear();
ct.teamid = result.teamid;
ct.appid = result.appid;
ct.expavg_time = dtime();
retval = ct.insert();
if (retval) return retval;
}
update_average(
now,
result.sent_time,
credit, CREDIT_HALF_LIFE,
ct.expavg, ct.expavg_time
);
sprintf(clause1,
"total=total+%.15e, expavg=%.15e, expavg_time=%.15e, njobs=njobs+1",
credit, ct.expavg, ct.expavg_time
);
sprintf(clause2,
"teamid=%lu and appid=%lu", result.teamid, result.appid
);
retval = ct.update_fields_noid(clause1, clause2);
if (retval) return retval;
return 0;
}
///////////////////// V2 CREDIT STUFF STARTS HERE ///////////////////
// levels of confidence in a credit value
//
#define PFC_MODE_NORMAL 0
// PFC was computed in the "normal" way, i.e.
// - claimed PFC
// - app version scaling (i.e. not anonymous platform)
// - host scaling
#define PFC_MODE_APPROX 1
// PFC was approximated, but still (in the absence of cheating)
// reflects the size of the particular job
#define PFC_MODE_WU_EST 2
// PFC was set to the WU estimate.
// If this doesn't reflect the WU size, neither does the PFC estimate
// This is a last resort, and can be way off.
#define PFC_MODE_INVALID 3
// PFC exceeded max granted credit - ignore
// used in the computation of AV scale factors
//
struct RSC_INFO {
double pfc_sum;
double pfc_n;
int nvers_thresh; // # app versions w/ lots of samples
int nvers_total;
RSC_INFO() {
pfc_sum = 0;
pfc_n = 0;
nvers_thresh = 0;
nvers_total = 0;
}
void update(APP_VERSION &av) {
nvers_total++;
if (av.pfc.n > MIN_VERSION_SAMPLES) {
nvers_thresh++;
pfc_sum += av.pfc.get_avg() * av.pfc.n;
pfc_n += av.pfc.n;
}
}
double avg() {
return pfc_sum/pfc_n;
}
};
// "avg" is the average PFC for this app
// over CPU versions or GPU versions, whichever is lowest.
// Update the pfc_scale of this app's versions in the DB,
// and update app.min_avg_pfc
//
int scale_versions(APP &app, double avg, SCHED_SHMEM *ssp) {
char buf[256];
int retval;
for (int j=0; j<ssp->napp_versions; j++) {
APP_VERSION &av = ssp->app_versions[j];
if (av.appid != app.id) {
continue;
}
if (av.pfc.n < MIN_VERSION_SAMPLES) {
continue;
}
av.pfc_scale= avg/av.pfc.get_avg();
DB_APP_VERSION dav;
dav.id = av.id;
sprintf(buf, "pfc_scale=%.15e", av.pfc_scale);
retval = dav.update_field(buf);
if (retval) {
return retval;
}
if (config.debug_credit) {
PLATFORM *p = ssp->lookup_platform_id(av.platformid);
log_messages.printf(MSG_NORMAL,
" updating scale factor for %lu (%s %s)\n",
av.id, p->name, av.plan_class
);
log_messages.printf(MSG_NORMAL,
" n: %g avg PFC: %g new scale: %g\n",
av.pfc.n, av.pfc.get_avg(), av.pfc_scale
);
}
}
app.min_avg_pfc = avg;
DB_APP da;
da.id = app.id;
sprintf(buf, "min_avg_pfc=%.15e", avg);
retval = da.update_field(buf);
if (retval) {
return retval;
}
return 0;
}
#define DEFAULT_GPU_SCALE 0.1
// if there are no CPU versions to compare against,
// multiply pfc_scale of GPU versions by this.
// This reflects the average lower efficiency of GPU apps compared to CPU apps.
// The observed values from SETI@home and Milkyway are 0.05 and 0.08.
// We'll be a little generous and call it 0.1
// Update app version scale factors,
// and find the min average PFC for each app.
// Called periodically from the master feeder.
//
int update_av_scales(SCHED_SHMEM *ssp) {
int i, j, retval;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL, "-- updating app version scales --\n");
}
for (i=0; i<ssp->napps; i++) {
APP &app = ssp->apps[i];
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL, "app %s (%lu)\n", app.name, app.id);
}
RSC_INFO cpu_info, gpu_info;
// find the average PFC of CPU and GPU versions
for (j=0; j<ssp->napp_versions; j++) {
APP_VERSION &avr = ssp->app_versions[j];
if (avr.appid != app.id) {
continue;
}
DB_APP_VERSION av;
retval = av.lookup_id(avr.id);
if (retval) {
return retval;
}
avr = av; // update shared mem array
if (plan_class_to_proc_type(av.plan_class) != PROC_TYPE_CPU) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"add to gpu totals: (%lu %s) %g %g\n",
av.id, av.plan_class, av.pfc.n, av.pfc.get_avg()
);
}
gpu_info.update(av);
} else {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"add to cpu totals: (%lu %s) %g %g\n",
av.id, av.plan_class, av.pfc.n, av.pfc.get_avg()
);
}
cpu_info.update(av);
}
}
// If there are both CPU and GPU versions,
// and at least 1 of each is above threshold,
// normalize to the min of the averages
//
// Otherwise, if either CPU or GPU has at least
// 2 versions above threshold, normalize to the average
//
if (cpu_info.nvers_thresh && gpu_info.nvers_thresh) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"CPU avg: %g; GPU avg: %g\n",
cpu_info.avg(), gpu_info.avg()
);
}
scale_versions(app,
cpu_info.avg()<gpu_info.avg()?cpu_info.avg():gpu_info.avg(),
ssp
);
} else if (cpu_info.nvers_thresh > 1) {
log_messages.printf(MSG_NORMAL,
"CPU avg: %g\n", cpu_info.avg()
);
scale_versions(app, cpu_info.avg(), ssp);
} else if (gpu_info.nvers_thresh > 0) {
log_messages.printf(MSG_NORMAL,
"GPU avg: %g\n", gpu_info.avg()
);
scale_versions(app, gpu_info.avg()*DEFAULT_GPU_SCALE, ssp);
}
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL, "-------------\n");
}
return 0;
}
- server: change the following from per-host to per-(host, app version): - daily quota mechanism - reliable mechanism (accelerated retries) - "trusted" mechanism (adaptive replication) - scheduler: enforce host scale probation only for apps with host_scale_check set. - validator: do scale probation on invalid results (need this in addition to error and timeout cases) - feeder: update app version scales every 10 min, not 10 sec - back-end apps: support --foo as well as -foo for options Notes: - If you have, say, cuda, cuda23 and cuda_fermi plan classes, a host will have separate quotas for each one. That means it could error out on 100 jobs for cuda_fermi, and when its quota goes to zero, error out on 100 jobs for cuda23, etc. This is intentional; there may be cases where one version works but not the others. - host.error_rate and host.max_results_day are deprecated TODO: - the values in the app table for limits on jobs in progress etc. should override rather than config.xml. Implementation notes: scheduler: process_request(): read all host_app_versions for host at start; Compute "reliable" and "trusted" for each one. write modified records at end get_app_version(): add "reliable_only" arg; if set, use only reliable versions skip over-quota versions Multi-pass scheduling: if have at least one reliable version, do a pass for jobs that need reliable, and use only reliable versions. Then clear best_app_versions cache. Score-based scheduling: for need-reliable jobs, it will pick the fastest version, then give a score bonus if that version happens to be reliable. When get back a successful result from client: increase daily quota When get back an error result from client: impose scale probation decrease daily quota if not aborted Validator: when handling a WU, create a vector of HOST_APP_VERSION parallel to vector of RESULT. Pass it to assign_credit_set(). Make copies of originals so we can update only modified ones update HOST_APP_VERSION error rates Transitioner: decrease quota on timeout svn path=/trunk/boinc/; revision=21181
2010-04-15 03:13:56 +00:00
// look up HOST_APP_VERSION record; called from validator and transitioner.
2012-10-17 01:44:55 +00:00
// Normally the record will exist; if not:
// look for another HOST_APP_VERSION for the same
// (host, app, platform, plan class).
// If find one, used it and update its app_version_id.
// Otherwise create a new one.
//
// This means that when a new app version is released,
// it inherits the runtime and reliability statistics of the old version.
// This is not always ideal (the new version may be faster/slower)
// but it's better than starting the statistics from scratch.
//
2012-10-17 17:37:51 +00:00
// (if anonymous platform, skip the above)
//
int hav_lookup(
DB_HOST_APP_VERSION &hav, DB_ID_TYPE hostid, DB_ID_TYPE gen_avid
) {
int retval;
char buf[256];
sprintf(buf, "where host_id=%lu and app_version_id=%ld", hostid, gen_avid);
retval = hav.lookup(buf);
2012-10-17 01:44:55 +00:00
if (retval != ERR_DB_NOT_FOUND) return retval;
2012-10-17 17:37:51 +00:00
// Here no HOST_APP_VERSION currently exists.
// If gen_avid is negative (anonymous platform) just make one
//
if (gen_avid < 0) {
hav.clear();
hav.host_id = hostid;
hav.app_version_id = gen_avid;
return hav.insert();
}
// otherwise try to appropriate an existing one as described above
//
2012-10-17 01:44:55 +00:00
DB_HOST_APP_VERSION hav2, best_hav;
DB_APP_VERSION av, av2, best_av;
2012-10-17 17:37:51 +00:00
retval = av.lookup_id(gen_avid);
2012-10-17 01:44:55 +00:00
if (retval) return retval;
2012-10-17 17:37:51 +00:00
// find the HOST_APP_VERSION w/ latest version num
// for this (app/platform/plan class) and appropriate it
//
2012-10-17 01:44:55 +00:00
bool found = false;
sprintf(buf, "where host_id=%lu", hostid);
2012-10-17 17:37:51 +00:00
while (1) {
retval = hav2.enumerate(buf);
if (retval == ERR_DB_NOT_FOUND) break;
if (retval) return retval;
2012-10-17 01:44:55 +00:00
retval = av2.lookup_id(hav2.app_version_id);
if (retval) continue;
if (av2.appid != av.appid) continue;
if (av2.platformid != av.platformid) continue;
if (strcmp(av2.plan_class, av.plan_class)) continue;
if (found) {
if (av2.version_num > best_av.version_num) {
best_av = av2;
best_hav = hav2;
}
} else {
found = true;
best_av = av2;
best_hav = hav2;
}
}
if (found) {
hav = best_hav;
char query[256], where_clause[256];
sprintf(query, "app_version_id=%ld", gen_avid);
2012-10-17 01:44:55 +00:00
sprintf(where_clause,
"host_id=%lu and app_version_id=%ld",
2012-10-17 01:44:55 +00:00
hostid, best_av.id
);
retval = hav.update_fields_noid(query, where_clause);
if (retval) return retval;
} else {
hav.clear();
hav.host_id = hostid;
2012-10-17 17:37:51 +00:00
hav.app_version_id = gen_avid;
retval = hav.insert();
2012-10-17 01:44:55 +00:00
if (retval) return retval;
}
2012-10-17 01:44:55 +00:00
return 0;
}
DB_APP_VERSION_VAL *av_lookup(
DB_ID_TYPE id, vector<DB_APP_VERSION_VAL>& app_versions
) {
for (unsigned int i=0; i<app_versions.size(); i++) {
if (app_versions[i].id == id) {
return &app_versions[i];
}
}
DB_APP_VERSION_VAL av;
int retval = av.lookup_id(id);
if (retval) {
return NULL;
}
app_versions.push_back(av);
return &(app_versions[app_versions.size()-1]);
}
// the estimated PFC for a given WU, in the absence of any other info
//
inline double wu_estimated_pfc(WORKUNIT &wu, DB_APP &app) {
double x = wu.rsc_fpops_est;
if (app.min_avg_pfc) {
x *= app.min_avg_pfc;
}
return x;
}
inline double wu_estimated_credit(WORKUNIT &wu, DB_APP &app) {
return wu_estimated_pfc(wu, app)*COBBLESTONE_SCALE;
}
inline bool is_pfc_sane(double x, WORKUNIT &wu, DB_APP &app) {
if (x > 1e4 || x < 1e-4) {
log_messages.printf(MSG_CRITICAL,
"Bad FLOP ratio (%f): check workunit.rsc_fpops_est for %s (app %s)\n",
x, wu.name, app.name
);
return false;
}
return true;
}
// Compute or estimate "claimed peak FLOP count" for a completed job.
// Possibly update host_app_version records and write to DB.
// Possibly update app_version records in memory and let caller write to DB,
// to merge DB writes
//
int get_pfc(
RESULT &r, WORKUNIT &wu, DB_APP &app, // in
vector<DB_APP_VERSION_VAL>&app_versions, // in/out
DB_HOST_APP_VERSION &hav, // in/out
double &pfc, // out
int &mode // out
){
DB_APP_VERSION_VAL *avp=0;
mode = PFC_MODE_APPROX;
if (r.runtime_outlier) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] runtime outlier, not updating stats\n",
r.id
);
}
}
// is result from old scheduler that didn't set r.app_version_id?
// if so, use WU estimate (this is a transient condition
// while projects upgrade server software)
//
if (r.app_version_id == 0) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] missing app_version_id (%ld): returning WU default %.2f\n",
r.id, r.app_version_id, wu_estimated_credit(wu, app)
);
}
mode = PFC_MODE_WU_EST;
pfc = wu_estimated_pfc(wu, app);
return 0;
}
// temporary kludge for SETI@home:
// if GPU initialization fails the app falls back to CPU.
//
if (strstr(r.stderr_out, "Device Emulation (CPU)")) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu][AV#%ld] CUDA app fell back to CPU; returning WU default %.2f\n",
r.id, r.app_version_id, wu.rsc_fpops_est*COBBLESTONE_SCALE
);
}
mode = PFC_MODE_WU_EST;
pfc = wu_estimated_pfc(wu, app);
return 0;
}
// transition case: there's no host_app_version record
//
- server: change the following from per-host to per-(host, app version): - daily quota mechanism - reliable mechanism (accelerated retries) - "trusted" mechanism (adaptive replication) - scheduler: enforce host scale probation only for apps with host_scale_check set. - validator: do scale probation on invalid results (need this in addition to error and timeout cases) - feeder: update app version scales every 10 min, not 10 sec - back-end apps: support --foo as well as -foo for options Notes: - If you have, say, cuda, cuda23 and cuda_fermi plan classes, a host will have separate quotas for each one. That means it could error out on 100 jobs for cuda_fermi, and when its quota goes to zero, error out on 100 jobs for cuda23, etc. This is intentional; there may be cases where one version works but not the others. - host.error_rate and host.max_results_day are deprecated TODO: - the values in the app table for limits on jobs in progress etc. should override rather than config.xml. Implementation notes: scheduler: process_request(): read all host_app_versions for host at start; Compute "reliable" and "trusted" for each one. write modified records at end get_app_version(): add "reliable_only" arg; if set, use only reliable versions skip over-quota versions Multi-pass scheduling: if have at least one reliable version, do a pass for jobs that need reliable, and use only reliable versions. Then clear best_app_versions cache. Score-based scheduling: for need-reliable jobs, it will pick the fastest version, then give a score bonus if that version happens to be reliable. When get back a successful result from client: increase daily quota When get back an error result from client: impose scale probation decrease daily quota if not aborted Validator: when handling a WU, create a vector of HOST_APP_VERSION parallel to vector of RESULT. Pass it to assign_credit_set(). Make copies of originals so we can update only modified ones update HOST_APP_VERSION error rates Transitioner: decrease quota on timeout svn path=/trunk/boinc/; revision=21181
2010-04-15 03:13:56 +00:00
if (!hav.host_id) {
mode = PFC_MODE_WU_EST;
pfc = wu_estimated_pfc(wu, app);
return 0;
- server: change the following from per-host to per-(host, app version): - daily quota mechanism - reliable mechanism (accelerated retries) - "trusted" mechanism (adaptive replication) - scheduler: enforce host scale probation only for apps with host_scale_check set. - validator: do scale probation on invalid results (need this in addition to error and timeout cases) - feeder: update app version scales every 10 min, not 10 sec - back-end apps: support --foo as well as -foo for options Notes: - If you have, say, cuda, cuda23 and cuda_fermi plan classes, a host will have separate quotas for each one. That means it could error out on 100 jobs for cuda_fermi, and when its quota goes to zero, error out on 100 jobs for cuda23, etc. This is intentional; there may be cases where one version works but not the others. - host.error_rate and host.max_results_day are deprecated TODO: - the values in the app table for limits on jobs in progress etc. should override rather than config.xml. Implementation notes: scheduler: process_request(): read all host_app_versions for host at start; Compute "reliable" and "trusted" for each one. write modified records at end get_app_version(): add "reliable_only" arg; if set, use only reliable versions skip over-quota versions Multi-pass scheduling: if have at least one reliable version, do a pass for jobs that need reliable, and use only reliable versions. Then clear best_app_versions cache. Score-based scheduling: for need-reliable jobs, it will pick the fastest version, then give a score bonus if that version happens to be reliable. When get back a successful result from client: increase daily quota When get back an error result from client: impose scale probation decrease daily quota if not aborted Validator: when handling a WU, create a vector of HOST_APP_VERSION parallel to vector of RESULT. Pass it to assign_credit_set(). Make copies of originals so we can update only modified ones update HOST_APP_VERSION error rates Transitioner: decrease quota on timeout svn path=/trunk/boinc/; revision=21181
2010-04-15 03:13:56 +00:00
}
// old clients report CPU time but not elapsed time.
// Use HOST_APP_VERSION.et to track statistics of CPU time.
//
if (r.elapsed_time < 1e-6) {
// in case buggy client reports elapsed time like 1e-304
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] old client (elapsed time not reported)\n",
r.id
);
}
if (!r.runtime_outlier) {
hav.et.update_var(
r.cpu_time/wu.rsc_fpops_est,
HAV_AVG_THRESH, HAV_AVG_WEIGHT, HAV_AVG_LIMIT
);
// if ((r.elapsed_time > 0) && (r.cpu_time > 0)) {
// hav.rt.update(r.elapsed_time,HAV_AVG_THRESH,HAV_AVG_WEIGHT,HAV_AVG_LIMIT);
// hav.cpu.update(r.cpu_time,HAV_AVG_THRESH,HAV_AVG_WEIGHT,HAV_AVG_LIMIT);
// }
}
pfc = wu_estimated_pfc(wu, app);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] old client: raw credit %.2f\n",
r.id, pfc*COBBLESTONE_SCALE
);
}
bool do_scale = true;
if (hav.et.n < MIN_HOST_SAMPLES || (hav.et.get_avg() <= 0)) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] old client: no host scaling - zero or too few samples %f\n",
r.id, hav.et.n
);
}
}
if (do_scale
&& app.host_scale_check
&& hav.consecutive_valid < CONS_VALID_HOST_SCALE
) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] old client: no host scaling - cons valid %d\n",
r.id, hav.consecutive_valid
);
}
}
if (do_scale) {
double s = r.cpu_time / (hav.et.get_avg()*wu.rsc_fpops_est);
pfc *= s;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] old client: scaling (based on CPU time) by %g, return %.2f\n",
r.id, s, pfc*COBBLESTONE_SCALE
);
}
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] old client: returning PFC %.2f\n",
r.id, pfc*COBBLESTONE_SCALE
);
}
return 0;
}
// r.flops_estimate should be positive
// but (because of scheduler bug) it may not be.
// At this point we don't have much to go on, so use 1e10.
//
if (r.flops_estimate <= 0) {
r.flops_estimate = 1e10;
}
double raw_pfc = (r.elapsed_time * r.flops_estimate);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] raw credit: %.2f (%.2f sec, %.2f est GFLOPS)\n",
r.id, raw_pfc*COBBLESTONE_SCALE, r.elapsed_time,
r.flops_estimate/1e9
);
}
// get app version
//
avp = av_lookup(r.app_version_id, app_versions);
// Sanity check
// If an app version scale exists, use it. Otherwise assume 1.
//
double tmp_scale = (avp && avp->pfc_scale) ? (avp->pfc_scale) : 1.0;
if (raw_pfc*tmp_scale > wu.rsc_fpops_bound) {
// This sanity check should be unnecessary because we have a maximum
// credit grant limit.
// With anonymous GPU apps the sanity check often fails
// because anonymous GPU scales are often of order 0.01.
// That prevents PFC averages from being updated.
// So I've removed the return statement.
//
pfc = wu_estimated_pfc(wu, app);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] WARNING: sanity check failed: %.2f>%.2f, return %.2f\n",
r.id, raw_pfc*tmp_scale*COBBLESTONE_SCALE,
wu.rsc_fpops_bound*COBBLESTONE_SCALE, pfc*COBBLESTONE_SCALE
);
}
// This was a bad idea because it prevents HAV.pfc from being updated.
// sprintf(query, "consecutive_valid=0");
// sprintf(clause, "host_id=%d and app_version_id=%d", r.hostid, gavid);
// retval = hav.update_fields_noid(query, clause);
// return retval;
}
if (r.app_version_id < 0) {
// anon platform
//
bool do_scale = true;
if (hav.pfc.n < MIN_HOST_SAMPLES || hav.pfc.get_avg()<=0) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] anon platform, not scaling, PFC avg zero or too few samples %.0f\n",
r.id, hav.pfc.n
);
}
}
if (do_scale
&& app.host_scale_check
&& hav.consecutive_valid < CONS_VALID_HOST_SCALE
) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] anon platform, not scaling, cons valid %d\n",
r.id, hav.consecutive_valid
);
}
}
if (do_scale) {
double scale = app.min_avg_pfc / hav.pfc.get_avg();
pfc = raw_pfc * scale;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] anon platform, scaling by %g (%.2f/%.2f)\n",
r.id, scale, app.min_avg_pfc, hav.pfc.get_avg()
);
}
} else {
pfc = wu_estimated_pfc(wu, app);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] not scaling, using app avg %.2f\n",
r.id, pfc*COBBLESTONE_SCALE
);
}
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] anon platform, returning %.2f\n",
r.id, pfc*COBBLESTONE_SCALE
);
}
} else {
avp = av_lookup(r.app_version_id, app_versions);
if (!avp) {
log_messages.printf(MSG_CRITICAL,
"get_pfc() [RESULT#%lu]: No AVP %ld!!\n", r.id, r.app_version_id
);
return ERR_NOT_FOUND;
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] [AV#%lu] normal case. %.0f sec, %.1f GFLOPS. raw credit: %.2f\n",
r.id, avp->id, r.elapsed_time, r.flops_estimate/1e9,
raw_pfc*COBBLESTONE_SCALE
);
}
bool do_scale = true;
double host_scale = 0;
if (app.host_scale_check
&& hav.consecutive_valid < CONS_VALID_HOST_SCALE
) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] not host scaling - cons valid %d\n",
r.id, hav.consecutive_valid
);
}
}
if (do_scale && (hav.pfc.n < MIN_HOST_SAMPLES || hav.pfc.get_avg()==0)) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] not host scaling - HAV PFC zero or too few samples %.0f\n",
r.id, hav.pfc.n
);
}
}
if (do_scale && avp->pfc.n < MIN_VERSION_SAMPLES) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] not host scaling - app_version PFC too few samples%.0f\n",
r.id, avp->pfc.n
);
}
}
if (do_scale && hav.pfc.get_avg() <= 0) {
do_scale = false;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] not host scaling - HAV PFC is zero\n",
r.id
);
}
}
if (do_scale) {
host_scale = avp->pfc.get_avg() / hav.pfc.get_avg();
if (host_scale > 10) {
host_scale = 10;
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] host scale: %.2f (%f/%f)\n",
r.id, host_scale, avp->pfc.get_avg(), hav.pfc.get_avg()
);
}
}
pfc = raw_pfc;
if (avp->pfc_scale) {
pfc *= avp->pfc_scale;
if (host_scale) {
pfc *= host_scale;
mode = PFC_MODE_NORMAL;
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] applying app version scale %.3f\n",
r.id, avp->pfc_scale
);
}
} else {
if (host_scale) {
pfc *= host_scale;
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] no app version scale\n",
r.id
);
}
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] [AV#%lu] PFC avgs with %g (%g/%g)\n",
r.id, avp->id,
raw_pfc/wu.rsc_fpops_est,
raw_pfc, wu.rsc_fpops_est
);
}
double x = raw_pfc / wu.rsc_fpops_est;
if (!r.runtime_outlier && is_pfc_sane(x, wu, app)) {
avp->pfc_samples.push_back(x);
}
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] updating HAV PFC %.2f et %g turnaround %d\n",
r.id, raw_pfc / wu.rsc_fpops_est,
r.elapsed_time / wu.rsc_fpops_est,
(r.received_time - r.sent_time)
);
}
if (!r.runtime_outlier) {
double x = raw_pfc / wu.rsc_fpops_est;
if (is_pfc_sane(x, wu, app)) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] [HOST#%lu] before updating HAV PFC pfc.n=%f pfc.avg=%f\n",
r.id, hav.host_id, hav.pfc.n, hav.pfc.avg
);
}
hav.pfc.update(x, HAV_AVG_THRESH, HAV_AVG_WEIGHT, HAV_AVG_LIMIT);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] [HOST#%lu] after updating HAV PFC pfc.n=%f pfc.avg=%f\n",
r.id, hav.host_id, hav.pfc.n, hav.pfc.avg
);
}
}
hav.et.update_var(
r.elapsed_time / wu.rsc_fpops_est,
HAV_AVG_THRESH, HAV_AVG_WEIGHT, HAV_AVG_LIMIT
);
hav.turnaround.update_var(
(r.received_time - r.sent_time),
HAV_AVG_THRESH, HAV_AVG_WEIGHT, HAV_AVG_LIMIT
);
}
// keep track of credit per app version
//
if (avp) {
avp->credit_samples.push_back(pfc*COBBLESTONE_SCALE);
avp->credit_times.push_back(r.sent_time);
}
return 0;
}
// compute the average of some numbers,
// where each value is weighted by the sum of the other values.
// (reduces the weight of large outliers)
//
double low_average(vector<double>& v) {
int i;
int n = v.size();
if (n == 1) {
return v[0];
}
double sum=0;
for (i=0; i<n; i++) {
sum += v[i];
}
double total=0;
for (i=0; i<n; i++) {
total += v[i]*(sum-v[i]);
}
return total/((n-1)*sum);
}
// compute the average of number weighted by proximity
// to another number
double pegged_average(vector<double>& v, double anchor) {
int n=v.size();
double weights=0,sum=0,w;
int i;
if (n==1) {
return v[0];
}
for (i=0; i<n; i++) {
w=(1.0/(0.1*anchor+fabs(anchor-v[i])));
weights+=w;
sum+=w*v[i];
}
return sum/weights;
}
double vec_min(vector<double>& v) {
double x = v[0];
for (unsigned int i=1; i<v.size(); i++) {
if (v[i] < x) {
x = v[i];
}
}
return x;
}
// Called by validator when canonical result has been selected.
// For each valid result in the list:
// - calculate a peak FLOP count (PFC) and a "mode" that indicates
// our confidence in the PFC
// - upate the statistics of PFC in host_app_version and app_version
// - Compute a credit value based on a weighted average of
// the PFCs of valid results
// (this value can be used or ignored by the caller)
//
// This must be called exactly once for each valid result.
//
int assign_credit_set(
WORKUNIT &wu,
vector<RESULT>& results,
DB_APP &app,
vector<DB_APP_VERSION_VAL>& app_versions,
- server: change the following from per-host to per-(host, app version): - daily quota mechanism - reliable mechanism (accelerated retries) - "trusted" mechanism (adaptive replication) - scheduler: enforce host scale probation only for apps with host_scale_check set. - validator: do scale probation on invalid results (need this in addition to error and timeout cases) - feeder: update app version scales every 10 min, not 10 sec - back-end apps: support --foo as well as -foo for options Notes: - If you have, say, cuda, cuda23 and cuda_fermi plan classes, a host will have separate quotas for each one. That means it could error out on 100 jobs for cuda_fermi, and when its quota goes to zero, error out on 100 jobs for cuda23, etc. This is intentional; there may be cases where one version works but not the others. - host.error_rate and host.max_results_day are deprecated TODO: - the values in the app table for limits on jobs in progress etc. should override rather than config.xml. Implementation notes: scheduler: process_request(): read all host_app_versions for host at start; Compute "reliable" and "trusted" for each one. write modified records at end get_app_version(): add "reliable_only" arg; if set, use only reliable versions skip over-quota versions Multi-pass scheduling: if have at least one reliable version, do a pass for jobs that need reliable, and use only reliable versions. Then clear best_app_versions cache. Score-based scheduling: for need-reliable jobs, it will pick the fastest version, then give a score bonus if that version happens to be reliable. When get back a successful result from client: increase daily quota When get back an error result from client: impose scale probation decrease daily quota if not aborted Validator: when handling a WU, create a vector of HOST_APP_VERSION parallel to vector of RESULT. Pass it to assign_credit_set(). Make copies of originals so we can update only modified ones update HOST_APP_VERSION error rates Transitioner: decrease quota on timeout svn path=/trunk/boinc/; revision=21181
2010-04-15 03:13:56 +00:00
vector<DB_HOST_APP_VERSION>& host_app_versions,
double max_granted_credit,
double &credit
) {
unsigned int i;
int mode, retval;
double pfc;
vector<double> normal;
vector<double> approx;
for (i=0; i<results.size(); i++) {
RESULT &r = results[i];
if (r.validate_state != VALIDATE_STATE_VALID) {
continue;
}
DB_HOST_APP_VERSION &hav = host_app_versions[i];
- server: change the following from per-host to per-(host, app version): - daily quota mechanism - reliable mechanism (accelerated retries) - "trusted" mechanism (adaptive replication) - scheduler: enforce host scale probation only for apps with host_scale_check set. - validator: do scale probation on invalid results (need this in addition to error and timeout cases) - feeder: update app version scales every 10 min, not 10 sec - back-end apps: support --foo as well as -foo for options Notes: - If you have, say, cuda, cuda23 and cuda_fermi plan classes, a host will have separate quotas for each one. That means it could error out on 100 jobs for cuda_fermi, and when its quota goes to zero, error out on 100 jobs for cuda23, etc. This is intentional; there may be cases where one version works but not the others. - host.error_rate and host.max_results_day are deprecated TODO: - the values in the app table for limits on jobs in progress etc. should override rather than config.xml. Implementation notes: scheduler: process_request(): read all host_app_versions for host at start; Compute "reliable" and "trusted" for each one. write modified records at end get_app_version(): add "reliable_only" arg; if set, use only reliable versions skip over-quota versions Multi-pass scheduling: if have at least one reliable version, do a pass for jobs that need reliable, and use only reliable versions. Then clear best_app_versions cache. Score-based scheduling: for need-reliable jobs, it will pick the fastest version, then give a score bonus if that version happens to be reliable. When get back a successful result from client: increase daily quota When get back an error result from client: impose scale probation decrease daily quota if not aborted Validator: when handling a WU, create a vector of HOST_APP_VERSION parallel to vector of RESULT. Pass it to assign_credit_set(). Make copies of originals so we can update only modified ones update HOST_APP_VERSION error rates Transitioner: decrease quota on timeout svn path=/trunk/boinc/; revision=21181
2010-04-15 03:13:56 +00:00
retval = get_pfc(r, wu, app, app_versions, hav, pfc, mode);
if (retval) {
log_messages.printf(MSG_CRITICAL,
"get_pfc() error: %s\n", boincerror(retval)
);
continue;
} else {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%lu] get_pfc() returns credit %g mode %s\n",
r.id, pfc *COBBLESTONE_SCALE, (mode==PFC_MODE_NORMAL)?"normal":"approx"
);
}
}
if (pfc > wu.rsc_fpops_bound) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] PFC too high: %f\n", pfc*COBBLESTONE_SCALE
);
}
pfc = wu_estimated_pfc(wu, app);
}
// max_granted_credit trumps rsc_fpops_bound;
// the latter may be set absurdly high
//
if (max_granted_credit && pfc*COBBLESTONE_SCALE > max_granted_credit) {
log_messages.printf(MSG_CRITICAL,
"[credit] Credit too high: %f\n", pfc*COBBLESTONE_SCALE
);
pfc = max_granted_credit/COBBLESTONE_SCALE;
mode = PFC_MODE_INVALID;
}
switch (mode) {
case PFC_MODE_NORMAL:
normal.push_back(pfc);
break;
case PFC_MODE_INVALID:
break;
default:
approx.push_back(pfc);
break;
}
}
// averaging policy: if there is more than one normal result,
// use the "pegged average" of normal results.
// Otherwise use the pegged_average of all results
//
double x;
switch (normal.size()) {
case 1:
// normal has double the weight of approx
approx.push_back(normal[0]);
approx.push_back(normal[0]);
// fall through
case 0:
if (approx.size()) {
x = pegged_average(approx, wu_estimated_pfc(wu, app));
} else {
// there were only PFC_MODE_INVALID results, so we guess
x = wu_estimated_pfc(wu, app);
}
break;
default:
x = pegged_average(normal, wu_estimated_pfc(wu, app));
break;
}
x *= COBBLESTONE_SCALE;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [WU#%lu] assign_credit_set: credit %g\n",
wu.id, x
);
}
credit = x;
return 0;
}
// carefully write any app_version records that have changed;
// done at the end of every validator scan.
//
int write_modified_app_versions(vector<DB_APP_VERSION_VAL>& app_versions) {
unsigned int i, j;
int retval = 0;
double now = dtime();
for (i=0; i<app_versions.size(); i++) {
DB_APP_VERSION_VAL &av = app_versions[i];
if (av.pfc_samples.empty() && av.credit_samples.empty()) {
continue;
}
for (int k=0; k<10; k++) {
double pfc_n_orig = av.pfc.n;
double expavg_credit_orig = av.expavg_credit;
for (j=0; j<av.pfc_samples.size(); j++) {
av.pfc.update(
av.pfc_samples[j],
AV_AVG_THRESH, AV_AVG_WEIGHT, AV_AVG_LIMIT
);
}
for (j=0; j<av.credit_samples.size(); j++) {
update_average(
now,
av.credit_times[j], av.credit_samples[j], CREDIT_HALF_LIFE,
av.expavg_credit, av.expavg_time
);
}
char query[512], clause[512];
sprintf(query,
"pfc_n=%.15e, pfc_avg=%.15e, expavg_credit=%.15e, expavg_time=%f",
av.pfc.n,
av.pfc.avg,
av.expavg_credit,
av.expavg_time
);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] updating app version %lu:\n", av.id
);
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] pfc.n = %f, pfc.avg = %f, expavg_credit = %f, expavg_time=%f\n",
av.pfc.n,
av.pfc.avg,
av.expavg_credit,
av.expavg_time
);
}
// if pfc_scale has changed (from feeder) reread it
//
sprintf(clause,
"pfc_n=%.15e and abs(expavg_credit-%.15e)<1e-4 and abs(pfc_scale-%.15e)<1e-6",
pfc_n_orig, expavg_credit_orig, av.pfc_scale
);
retval = av.update_field(query, clause);
if (retval) {
break;
}
if (boinc_db.affected_rows() == 1) {
break;
}
retval = av.lookup_id(av.id);
if (retval) {
break;
}
}
av.pfc_samples.clear();
av.credit_samples.clear();
av.credit_times.clear();
if (retval) {
return retval;
}
}
return 0;
}