// Called by validator when canonical result has been selected.
// Compute credit for valid instances.
// This is called exactly once for each valid result.
//
int assign_credit_set(
WORKUNIT& wu, vector<RESULT>& results,
DB_APP& app,
vector<DB_APP_VERSION>& app_versions,
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];
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#%d] 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) {
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_NORMAL,
"[credit] Credit too high: %f\n", pfc*COBBLESTONE_SCALE
);
pfc = max_granted_credit/COBBLESTONE_SCALE;
}
if (mode == PFC_MODE_NORMAL) {
normal.push_back(pfc);
} else {
approx.push_back(pfc);
}
}
// averaging policy: if there is least one normal result,
// use the "low average" of normal results.
// Otherwise use the min of all results
//
double x;
if (normal.size()) {
x = low_average(normal);
} else if (approx.size()) {
x = vec_min(approx);
} else {
x = 0;
}
x *= COBBLESTONE_SCALE;
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [WU#%d] assign_credit_set: credit %g\n",
wu.id, x
);
}
credit = x;
return 0;
}
// Compute or estimate "claimed peak FLOP count".
// 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>&app_versions, // in/out
DB_HOST_APP_VERSION& hav, // in/out
double& pfc, int& mode // out
) {
DB_APP_VERSION* avp=0;
int retval;
mode = PFC_MODE_APPROX;
if (r.runtime_outlier && config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%d] runtime outlier, not updating stats\n",
r.id
);
}
// is result from old scheduler that didn't set r.app_version_id correctly?
// if so, use WU estimate (this is a transient condition)
//
if (r.app_version_id == 0 || r.app_version_id == 1) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%d] missing app_version_id (%d): 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#%d][AV#%d] 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;
}
int gavid = generalized_app_version_id(r.app_version_id, r.appid);
// transition case
//
if (!hav.host_id) {
mode = PFC_MODE_WU_EST;
pfc = wu_estimated_pfc(wu, app);
return 0;
}
// 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#%d] 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
);
}
pfc = wu_estimated_pfc(wu, app);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%d] 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#%d] 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#%d] 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#%d] 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#%d] 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#%d] raw credit: %.2f (%.2f sec, %.2f est GFLOPS)\n",
r.id, raw_pfc*COBBLESTONE_SCALE, r.elapsed_time,
r.flops_estimate/1e9
);
}
// Sanity check
//
if (raw_pfc > wu.rsc_fpops_bound) {
char query[256], clause[256];
pfc = wu_estimated_pfc(wu, app);
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%d] sanity check failed: %.2f>%.2f, return %.2f\n",
r.id, raw_pfc*COBBLESTONE_SCALE,
wu.rsc_fpops_bound*COBBLESTONE_SCALE, pfc*COBBLESTONE_SCALE
);
}
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#%d] 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#%d] 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#%d] 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#%d] not scaling, using app avg %.2f\n",
r.id, pfc*COBBLESTONE_SCALE
);
}
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%d] 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#%d]: No AVP %d!!\n", r.id, r.app_version_id
);
return ERR_NOT_FOUND;
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%d] [AV#%d] 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#%d] 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#%d] 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#%d] 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#%d] 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#%d] 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#%d] 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#%d] no app version scale\n",
r.id
);
}
}
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"[credit] [RESULT#%d] [AV#%d] 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#%d] 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)
);
}
double x = raw_pfc / wu.rsc_fpops_est;
if (!r.runtime_outlier && is_pfc_sane(x, wu, app)) {
hav.pfc.update(x, HAV_AVG_THRESH, HAV_AVG_WEIGHT, HAV_AVG_LIMIT);
}
if (!r.runtime_outlier) {
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;
}
// 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,
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];
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;
}
inline double wu_estimated_credit(WORKUNIT& wu, DB_APP& app) {
return wu_estimated_pfc(wu, app)*COBBLESTONE_SCALE;
}
// 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
//
if (!hav.host_id) {
mode = PFC_MODE_WU_EST;
pfc = wu_estimated_pfc(wu, app);
return 0;
}
// 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;
}
