// "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 %d (%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;
}
DB_APP_VERSION* av_lookup(int id, vector<DB_APP_VERSION>& 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 av;
int retval = av.lookup_id(id);
if (retval) return NULL;
app_versions.push_back(av);
return &(app_versions[app_versions.size()-1]);
}
// return total project FLOPS (based on recent credit)
//
int get_project_flops(double& total) {
DB_APP_VERSION av;
char buf[256];
// compute credit per day
//
sprintf(buf, "where expavg_time > %f", dtime() - 30*86400);
total = 0;
while (1) {
int retval = av.enumerate(buf);
if (retval == ERR_DB_NOT_FOUND) break;
if (retval) {
return retval;
}
total += av.expavg_credit;
}
total /= COBBLESTONE_SCALE; // convert to FLOPs per day
total /= 86400; // convert to FLOPs per second
return 0;
}
// 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 (%d)\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 (app_plan_uses_gpu(av.plan_class)) {
if (config.debug_credit) {
log_messages.printf(MSG_NORMAL,
"add to gpu totals: (%d %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: (%d %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;
}
// The WU is already committed to an app version.
// - check if this host supports that platform
// - if plan class, check if this host can handle it
// - check if we need work for the resource
//
// If all these are satisfied, return a pointer to a BEST_APP_VERSION struct
// with HOST_USAGE filled in correctly.
// Else return NULL.
//
static BEST_APP_VERSION* check_homogeneous_app_version(
WORKUNIT& wu, bool /* reliable_only */
// TODO: enforce reliable_only
) {
BEST_APP_VERSION bav;
bool found;
APP_VERSION *avp = ssp->lookup_app_version(wu.app_version_id);
if (!avp) {
// If the app version is not in shmem,
// it's been superceded or deprecated.
// Use it anyway.
// Keep an array of such app versions in
// SCHEDULER_REPLY::old_app_versions
//
found = false;
for (unsigned int i=0; i<g_reply->old_app_versions.size(); i++) {
APP_VERSION& av = g_reply->old_app_versions[i];
if (av.id == wu.app_version_id) {
avp = &av;
found = true;
break;
}
}
if (!found) {
DB_APP_VERSION av;
int retval = av.lookup_id(wu.app_version_id);
if (retval) return NULL;
g_reply->old_app_versions.push_back(av);
avp = &(g_reply->old_app_versions.back());
}
}
// see if this host supports the version's platform
//
found = false;
for (unsigned int i=0; i<g_request->platforms.list.size(); i++) {
PLATFORM* p = g_request->platforms.list[i];
if (p->id == avp->platformid) {
found = true;
bav.avp = avp;
break;
}
}
if (!found) return NULL;
// and see if it supports the plan class
//
if (strlen(avp->plan_class)) {
if (!app_plan(*g_request, avp->plan_class, bav.host_usage)) {
return NULL;
}
} else {
bav.host_usage.sequential_app(capped_host_fpops());
}
// and see if the client is asking for this resource
//
if (!need_this_resource(bav.host_usage, avp, NULL)) {
return NULL;
}
// dynamically allocate the BEST_APP_VERSION.
// This is a memory leak, but that's OK
//
BEST_APP_VERSION* bavp = new BEST_APP_VERSION;
*bavp = bav;
return bavp;
}