From 01739471e11114ef196bf7da7ccb852fbf922d43 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ma=C3=ABl=20Madon?= <mael.madon@irit.fr>
Date: Thu, 20 Jul 2023 18:20:48 +0200
Subject: [PATCH] fix: propagate bug fix easy_bf_fast, see batsched issue #9
---
src/scheds/easy_bf_fast.cpp | 51 ++++++++++++++++++++-----------------
src/scheds/easy_bf_fast.hpp | 4 ++-
2 files changed, 31 insertions(+), 24 deletions(-)
diff --git a/src/scheds/easy_bf_fast.cpp b/src/scheds/easy_bf_fast.cpp
index 1a9112a..4bc094e 100644
--- a/src/scheds/easy_bf_fast.cpp
+++ b/src/scheds/easy_bf_fast.cpp
@@ -52,8 +52,11 @@ void EasyBackfillingFast::make_decisions(double date,
// - only handles one priority job (the first of the queue)
// - only handles time as floating-point (-> precision errors).
- // Hacks:
- // - uses priority job's completion time to store its expected starting time
+ // Warning: you might obtain different outputs than with easy_bf. This is
+ // due to the fact that this version only keeps track of the priority job
+ // expected start time and the number of machines available then, while
+ // easy_bf keeps track of a full 2D schedule of the future. easy_bf_fast
+ // will sometimes be a little more greedy in backfilling.
bool job_ended = false;
@@ -134,8 +137,7 @@ void EasyBackfillingFast::make_decisions(double date,
{
// The job becomes priority!
_priority_job = pending_job;
- _priority_job->completion_time
- = compute_priority_job_expected_earliest_starting_time();
+ update_priority_job_expected_earliest_start_time();
_pending_jobs.erase(job_it);
// Stop first queue traversal.
@@ -150,17 +152,16 @@ void EasyBackfillingFast::make_decisions(double date,
// Update priority job expected starting time (might have changed if a recently ended job
// completed before its walltime)
if (_priority_job != nullptr)
- _priority_job->completion_time = compute_priority_job_expected_earliest_starting_time();
+ update_priority_job_expected_earliest_start_time();
for (auto job_it = _pending_jobs.begin();
job_it != _pending_jobs.end();)
{
const Job *pending_job = *job_it;
- // Can the job be executed now ?
- if (pending_job->nb_requested_resources
- <= _nb_available_machines
- && date + pending_job->walltime
- <= _priority_job->completion_time)
+ // Can the job be executed now (without hindering priority job)?
+ if (pending_job->nb_requested_resources <= _nb_available_machines &&
+ (date + pending_job->walltime <= _priority_job_expected_start_time ||
+ pending_job->nb_requested_resources <= _remaining_resources_at_priority_job_start))
{
// Yes, it can be backfilled!
alloc.machines = _available_machines.left(
@@ -182,6 +183,8 @@ void EasyBackfillingFast::make_decisions(double date,
-= pending_job->nb_requested_resources;
_current_allocations[pending_job->id] = alloc;
job_it = _pending_jobs.erase(job_it);
+ if(date + pending_job->walltime > _priority_job_expected_start_time)
+ _remaining_resources_at_priority_job_start -= pending_job->nb_requested_resources;
// Directly get out of the backfilling loop if all
// machines are busy.
@@ -222,13 +225,13 @@ void EasyBackfillingFast::make_decisions(double date,
{
// LOG_F(INFO, "There are enough available resources (%d) to execute
// job %s", _nb_available_machines, new_job->id.c_str());
- // Can it be executed now (without hindering priority job?)
- if (_priority_job == nullptr
- || date + new_job->walltime <= _priority_job->completion_time)
+ // Can it be executed now (without hindering priority job)?
+ if (_priority_job == nullptr ||
+ date + new_job->walltime <= _priority_job_expected_start_time ||
+ new_job->nb_requested_resources <= _remaining_resources_at_priority_job_start)
{
- // LOG_F(INFO, "Job %s can be started right away!",
- // new_job->id.c_str());
- // Yes, the job can be executed right away!
+ //LOG_F(INFO, "Job %s can be started right away!", new_job->id.c_str());
+ // Yes, the job can be executed right away!
Allocation alloc;
alloc.machines
@@ -246,6 +249,8 @@ void EasyBackfillingFast::make_decisions(double date,
_available_machines -= alloc.machines;
_nb_available_machines -= new_job->nb_requested_resources;
_current_allocations[new_job_id] = alloc;
+ if(_priority_job != nullptr && date + new_job->walltime > _priority_job_expected_start_time)
+ _remaining_resources_at_priority_job_start -= new_job->nb_requested_resources;
}
else
{
@@ -253,7 +258,7 @@ void EasyBackfillingFast::make_decisions(double date,
/*LOG_F(INFO, "Not enough time to execute job %s (walltime=%g,
priority job expected starting time=%g)",
new_job->id.c_str(), (double)new_job->walltime,
- _priority_job->completion_time);*/
+ _priority_job_expected_start_time);*/
_pending_jobs.push_back(new_job);
}
}
@@ -264,8 +269,7 @@ void EasyBackfillingFast::make_decisions(double date,
{
// The job becomes priority.
_priority_job = new_job;
- _priority_job->completion_time
- = compute_priority_job_expected_earliest_starting_time();
+ update_priority_job_expected_earliest_start_time();
}
else
{
@@ -280,8 +284,7 @@ void EasyBackfillingFast::make_decisions(double date,
DynScheduler::make_decisions(date, update_info, compare_info);
}
-double
-EasyBackfillingFast::compute_priority_job_expected_earliest_starting_time()
+void EasyBackfillingFast::update_priority_job_expected_earliest_start_time()
{
int nb_available = _nb_available_machines;
int required = _priority_job->nb_requested_resources;
@@ -292,12 +295,14 @@ EasyBackfillingFast::compute_priority_job_expected_earliest_starting_time()
if (nb_available >= required)
{
- return it->date;
+ _priority_job_expected_start_time = it->date;
+ _remaining_resources_at_priority_job_start = nb_available - required;
+ return;
}
}
PPK_ASSERT_ERROR(false, "The job will never be executable.");
- return 0;
+ return;
}
std::list<EasyBackfillingFast::FinishedHorizonPoint>::iterator
diff --git a/src/scheds/easy_bf_fast.hpp b/src/scheds/easy_bf_fast.hpp
index 270d720..b98b6cd 100644
--- a/src/scheds/easy_bf_fast.hpp
+++ b/src/scheds/easy_bf_fast.hpp
@@ -38,7 +38,7 @@ private:
};
private:
- double compute_priority_job_expected_earliest_starting_time();
+ void update_priority_job_expected_earliest_start_time();
std::list<FinishedHorizonPoint>::iterator insert_horizon_point(const FinishedHorizonPoint & point);
private:
@@ -58,4 +58,6 @@ private:
// At any time, null if there is no priority job (no waiting job)
Job * _priority_job = nullptr;
+ double _priority_job_expected_start_time = -1;
+ int _remaining_resources_at_priority_job_start = -1;
};
--
GitLab