boinc/sched/sched_score.cpp

// 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/>.

// Matchmaker scheduling code

#include <algorithm>

#include "boinc_db.h"
#include "error_numbers.h"
#include "util.h"

#include "sched_check.h"
#include "sched_config.h"
#include "sched_hr.h"
#include "sched_main.h"
#include "sched_msgs.h"
#include "sched_send.h"
#include "sched_shmem.h"
#include "sched_types.h"
#include "sched_version.h"

#include "sched_score.h"

#ifdef NEW_SCORE

static int get_size_class(APP& app, double es) {
    for (int i=0; i<app.n_size_classes-1; i++) {
        if (es < app.size_class_quantiles[i]) return i;
    }
    return app.n_size_classes - 1;
}

// Assign a score to this job,
// representing the value of sending the job to this host.
// Also do some initial screening,
// and return false if can't send the job to host
//
bool JOB::get_score(WU_RESULT& wu_result) {
    score = 0;

    if (!app->beta && wu_result.need_reliable) {
        if (!bavp->reliable) {
            return false;
        }
    }

    if (app->beta) {
        if (g_wreq->allow_beta_work) {
            score += 1;
        } else {
            return false;
        }
    }

    if (app_not_selected(wu_result.workunit)) {
        if (g_wreq->allow_non_preferred_apps) {
            score -= 1;
        } else {
            return false;
        }
    }

    if (wu_result.infeasible_count) {
        score += 1;
    }

    if (app->locality_scheduling == LOCALITY_SCHED_LITE
        && g_request->file_infos.size()
    ) {
        int n = nfiles_on_host(wu_result.workunit);
        if (config.debug_locality_lite) {
            log_messages.printf(MSG_NORMAL,
                "[loc_lite] job %s has %d files on this host\n",
                wu_result.workunit.name, n
            );
        }
        if (n > 0) {
            score += 10;
        }
    }

    if (app->n_size_classes > 1) {
        double effective_speed = bavp->host_usage.projected_flops * g_reply->host.on_frac * g_reply->host.active_frac;
        int target_size = get_size_class(*app, effective_speed);
        if (config.debug_send) {
            log_messages.printf(MSG_NORMAL,
                "[send] size: host %d job %d speed %f\n",
                target_size, wu_result.workunit.size_class, effective_speed
            );
        }
        if (target_size == wu_result.workunit.size_class) {
            score += 5;
        } else if (target_size < wu_result.workunit.size_class) {
            score -= 2;
        } else {
            score -= 1;
        }
    }
    if (config.debug_send) {
        log_messages.printf(MSG_NORMAL,
            "[send]: job score %f\n", score
        );
    }

    return true;
}

bool job_compare(JOB j1, JOB j2) {
    return (j1.score > j2.score);
}

static double req_sec_save[NPROC_TYPES];
static double req_inst_save[NPROC_TYPES];

static void clear_others(int rt) {
    for (int i=0; i<NPROC_TYPES; i++) {
        if (i == rt) continue;
        req_sec_save[i] = g_wreq->req_secs[i];
        g_wreq->req_secs[i] = 0;
        req_inst_save[i] = g_wreq->req_instances[i];
        g_wreq->req_instances[i] = 0;
    }
}

static void restore_others(int rt) {
    for (int i=0; i<NPROC_TYPES; i++) {
        if (i == rt) continue;
        g_wreq->req_secs[i] += req_sec_save[i];
        g_wreq->req_instances[i] += req_inst_save[i];
    }
}

// send work for a particular processor type
//
void send_work_score_type(int rt) {
    vector<JOB> jobs;

    if (config.debug_send) {
        log_messages.printf(MSG_NORMAL,
            "[send] scanning for %s jobs\n", proc_type_name(rt)
        );
    }

    clear_others(rt);

    int nscan = config.mm_max_slots;
    if (!nscan) nscan = ssp->max_wu_results;
    int rnd_off = rand() % ssp->max_wu_results;
    for (int j=0; j<nscan; j++) {
        int i = (j+rnd_off) % ssp->max_wu_results;
        WU_RESULT& wu_result = ssp->wu_results[i];
        if (wu_result.state != WR_STATE_PRESENT) {
            continue;
        }
        WORKUNIT wu = wu_result.workunit;
        JOB job;
        job.app = ssp->lookup_app(wu.appid);
        if (job.app->non_cpu_intensive) continue;
        job.bavp = get_app_version(wu, true, false);
        if (!job.bavp) continue;

        job.index = i;
        job.result_id = wu_result.resultid;
        if (!job.get_score(wu_result)) {
            continue;
        }
        jobs.push_back(job);
    }

    std::sort(jobs.begin(), jobs.end(), job_compare);

    bool sema_locked = false;
    for (unsigned int i=0; i<jobs.size(); i++) {
        if (!work_needed(false)) {
            break;
        }
        if (!g_wreq->need_proc_type(rt)) {
            break;
        }
        JOB& job = jobs[i];
        if (!sema_locked) {
            lock_sema();
            sema_locked = true;
        }

        // make sure the job is still in the cache
        // array is locked at this point.
        //
        WU_RESULT& wu_result = ssp->wu_results[job.index];
        if (wu_result.state != WR_STATE_PRESENT) {
            continue;
        }
        if (wu_result.resultid != job.result_id) {
            continue;
        }
        WORKUNIT wu = wu_result.workunit;
        int retval = wu_is_infeasible_fast(
            wu,
            wu_result.res_server_state, wu_result.res_priority,
            wu_result.res_report_deadline,
            *job.app,
            *job.bavp
        );

        if (retval) {
            continue;
        }
        wu_result.state = g_pid;

        // It passed fast checks.
        // Release sema and to slow checks
        unlock_sema();
        sema_locked = false;

        switch (slow_check(wu_result, job.app, job.bavp)) {
        case 1:
            wu_result.state = WR_STATE_PRESENT;
            break;
        case 2:
            wu_result.state = WR_STATE_EMPTY;
            break;
        default:
            // slow_check() refreshes fields of wu_result.workunit;
            // update our copy too
            //
            wu.hr_class = wu_result.workunit.hr_class;
            wu.app_version_id = wu_result.workunit.app_version_id;

            // mark slot as empty AFTER we've copied out of it
            // (since otherwise feeder might overwrite it)
            //
            wu_result.state = WR_STATE_EMPTY;

            // reread result from DB, make sure it's still unsent
            // TODO: from here to end of add_result_to_reply()
            // (which updates the DB record) should be a transaction
            //
            SCHED_DB_RESULT result;
            result.id = wu_result.resultid;
            if (result_still_sendable(result, wu)) {
                add_result_to_reply(result, wu, job.bavp, false);

                // add_result_to_reply() fails only in pathological cases -
                // e.g. we couldn't update the DB record or modify XML fields.
                // If this happens, don't replace the record in the array
                // (we can't anyway, since we marked the entry as "empty").
                // The feeder will eventually pick it up again,
                // and hopefully the problem won't happen twice.
            }
            break;
        }
    }
    if (sema_locked) {
        unlock_sema();
    }

    restore_others(rt);
    g_wreq->best_app_versions.clear();
}

void send_work_score() {
    for (int i=0; i<NPROC_TYPES; i++) {
        if (g_wreq->need_proc_type(i)) {
            send_work_score_type(i);
        }
    }
}
#else
// reread result from DB, make sure it's still unsent
// TODO: from here to add_result_to_reply()
// (which updates the DB record) should be a transaction
//
int read_sendable_result(SCHED_DB_RESULT& result) {
    int retval = result.lookup_id(result.id);
    if (retval) {
        log_messages.printf(MSG_CRITICAL,
            "[RESULT#%u] result.lookup_id() failed %s\n",
            result.id, boincerror(retval)
        );
        return ERR_NOT_FOUND;
    }
    if (result.server_state != RESULT_SERVER_STATE_UNSENT) {
        log_messages.printf(MSG_NORMAL,
            "[RESULT#%u] expected to be unsent; instead, state is %d\n",
            result.id, result.server_state
        );
        return ERR_BAD_RESULT_STATE;
    }
    return 0;
}

// TODO: use slow_check()
//
bool wu_is_infeasible_slow(
    WU_RESULT& wu_result, SCHEDULER_REQUEST&, SCHEDULER_REPLY&
) {
    char buf[256];
    int retval;
    int n;
    SCHED_DB_RESULT result;

    // Don't send if we've already sent a result of this WU to this user.
    //
    if (config.one_result_per_user_per_wu) {
        sprintf(buf,
            "where workunitid=%d and userid=%d",
            wu_result.workunit.id, g_reply->user.id
        );
        retval = result.count(n, buf);
        if (retval) {
            log_messages.printf(MSG_CRITICAL,
                "send_work: can't get result count (%s)\n", boincerror(retval)
            );
            return true;
        } else {
            if (n>0) {
                if (config.debug_send) {
                    log_messages.printf(MSG_NORMAL,
                        "[send] send_work: user %d already has %d result(s) for WU %d\n",
                        g_reply->user.id, n, wu_result.workunit.id
                    );
                }
                return true;
            }
        }
    } else if (config.one_result_per_host_per_wu) {
        // Don't send if we've already sent a result
        // of this WU to this host.
        // We only have to check this
        // if we don't send one result per user.
        //
        sprintf(buf,
            "where workunitid=%d and hostid=%d",
            wu_result.workunit.id, g_reply->host.id
        );
        retval = result.count(n, buf);
        if (retval) {
            log_messages.printf(MSG_CRITICAL,
                "send_work: can't get result count (%s)\n", boincerror(retval)
            );
            return true;
        } else {
            if (n>0) {
                if (config.debug_send) {
                    log_messages.printf(MSG_NORMAL,
                        "[send] send_work: host %d already has %d result(s) for WU %d\n",
                        g_reply->host.id, n, wu_result.workunit.id
                    );
                }
                return true;
            }
        }
    }

    APP* app = ssp->lookup_app(wu_result.workunit.appid);
    WORKUNIT wu = wu_result.workunit;
    if (app_hr_type(*app)) {
        if (already_sent_to_different_hr_class(wu, *app)) {
            if (config.debug_send) {
                log_messages.printf(MSG_NORMAL,
                    "[send] [HOST#%d] [WU#%u %s] WU is infeasible (assigned to different platform)\n",
                    g_reply->host.id, wu.id, wu.name
                );
            }
            // Mark the workunit as infeasible.
            // This ensures that jobs already assigned to a platform
            // are processed first.
            //
            wu_result.infeasible_count++;
            return true;
        }
    }
    return false;
}

double JOB_SET::lowest_score() {
    if (jobs.empty()) return 0;
    return jobs.back().score;
}

// add the given job, and remove lowest-score jobs that
// - are in excess of work request
// - are in excess of per-request or per-day limits
// - cause the disk limit to be exceeded
//
void JOB_SET::add_job(JOB& job) {
    while (!jobs.empty()) {
        JOB& worst_job = jobs.back();
        if (est_time + job.est_time - worst_job.est_time > work_req) {
            est_time -= worst_job.est_time;
            disk_usage -= worst_job.disk_usage;
            jobs.pop_back();
            ssp->wu_results[worst_job.index].state = WR_STATE_PRESENT;
        } else {
            break;
        }
    }
    while (!jobs.empty()) {
        JOB& worst_job = jobs.back();
        if (disk_usage + job.disk_usage > disk_limit) {
            est_time -= worst_job.est_time;
            disk_usage -= worst_job.disk_usage;
            jobs.pop_back();
            ssp->wu_results[worst_job.index].state = WR_STATE_PRESENT;
        } else {
            break;
        }
    }

    if ((int)jobs.size() == max_jobs) {
        JOB& worst_job = jobs.back();
        jobs.pop_back();
        ssp->wu_results[worst_job.index].state = WR_STATE_PRESENT;
    }

    std::list<JOB>::iterator i = jobs.begin();
    while (i != jobs.end()) {
        if (i->score < job.score) {
            jobs.insert(i, job);
            break;
        }
        i++;
    }
    if (i == jobs.end()) {
        jobs.push_back(job);
    }
    est_time += job.est_time;
    disk_usage += job.disk_usage;
    if (config.debug_send) {
        log_messages.printf(MSG_NORMAL,
            "[send] added job to set.  est_time %.2f disk_usage %.2fGB\n",
            est_time, disk_usage/GIGA
        );
    }
}

// return the disk usage of jobs above the given score
//
double JOB_SET::higher_score_disk_usage(double v) {
    double sum = 0;
    std::list<JOB>::iterator i = jobs.begin();
    while (i != jobs.end()) {
        if (i->score < v) break;
        sum += i->disk_usage;
        i++;
    }
    return sum;
}

void JOB_SET::send() {
    WORKUNIT wu;
    SCHED_DB_RESULT result;
    int retval;

    std::list<JOB>::iterator i = jobs.begin();
    while (i != jobs.end()) {
        JOB& job = *(i++);
        WU_RESULT wu_result = ssp->wu_results[job.index];
        ssp->wu_results[job.index].state = WR_STATE_EMPTY;
        wu = wu_result.workunit;
        result.id = wu_result.resultid;
        retval = read_sendable_result(result);
        if (!retval) {
            add_result_to_reply(result, wu, job.bavp, false);
        }
    }
}

void send_work_score() {
    int i, slots_locked=0, slots_nonempty=0;
    JOB_SET jobs;
    int min_slots = config.mm_min_slots;
    if (!min_slots) min_slots = ssp->max_wu_results/2;
    int max_slots = config.mm_max_slots;
    if (!max_slots) max_slots = ssp->max_wu_results;
    int max_locked = 10;

    lock_sema();
    i = rand() % ssp->max_wu_results;

    // scan through the job cache, maintaining a JOB_SET of jobs
    // that we can send to this client, ordered by score.
    //
    for (int slots_scanned=0; slots_scanned<max_slots; slots_scanned++) {
        i = (i+1) % ssp->max_wu_results;
        WU_RESULT& wu_result = ssp->wu_results[i];
        switch (wu_result.state) {
        case WR_STATE_EMPTY:
            continue;
        case WR_STATE_PRESENT:
            slots_nonempty++;
            break;
        default:
            slots_nonempty++;
            if (wu_result.state == g_pid) break;
            slots_locked++;
            continue;
        }

        JOB job;
        job.index = i;

        // get score for this job, and skip it if it fails quick check.
        // NOTE: the EDF check done in get_score()
        // includes only in-progress jobs.
        //
        if (!job.get_score()) {
            continue;
        }
        if (config.debug_send) {
            log_messages.printf(MSG_NORMAL,
                "[send] score for %s: %f\n", wu_result.workunit.name, job.score
            );
        }

        if (job.score > jobs.lowest_score() || !jobs.request_satisfied()) {
            ssp->wu_results[i].state = g_pid;
            unlock_sema();
            if (wu_is_infeasible_slow(wu_result, *g_request, *g_reply)) {
                // if we can't use this job, put it back in pool
                //
                lock_sema();
                ssp->wu_results[i].state = WR_STATE_PRESENT;
                continue;
            }
            lock_sema();
            jobs.add_job(job);
        }

        if (jobs.request_satisfied() && slots_scanned>=min_slots) break;
    }

    if (slots_nonempty) {
        g_wreq->no_jobs_available = false;
    } else {
        log_messages.printf(MSG_CRITICAL,
            "Job cache is empty - check feeder\n"
        );
    }

    // TODO: trim jobs from tail of list until we pass the EDF check
    //
    jobs.send();
    unlock_sema();
    if (slots_locked > max_locked) {
        log_messages.printf(MSG_CRITICAL,
            "Found too many locked slots (%d>%d) - increase array size\n",
            slots_locked, max_locked
        );
    }
}

#endif