wiki:ClientSched

Version 5 (modified by davea, 16 years ago) (diff)

--

Client scheduling policies

NOTE: this document is outdated because of development in the 6.4 client to handle GPU and multithread apps. Changes include:

0) app versions now include avg_ncpus, coprocessor usage, and a FLOPS estimate (this defaults to the CPU benchmark, but for GPU and multithread apps it will be different). This info is sent from the server.

1) Estimating the duration of unstarted jobs: jobs are now associated with specific app versions. The estimated duration of an unstarted job is the WU's FLOP estimate divided by the app version's FLOPS, scaled by the duration correction factor.

2) Duration correction factor: this is now based on elapsed time (i.e. wall time during which the job has been running) rather than CPU time. Yes, this is affected by non-BOINC CPU load; that's as it should be.

3) "CPU efficiency" is no longer maintained; it's subsumed in DCF.

4) Estimating the duration of running jobs: this is a weighted average of static and dynamic estimates. The dynamic estimate is now based on elapsed time rather than CPU time. So if a GPU job has been running for 5 min, is 25% done, and has used 1 min of CPU, its dynamic estimate is 20 min (not 4 min).

5) round-robin simulation: this was modified to reflect multi-thread and coproc apps (e.g., if the host has 1 GPU, only one coproc app can run at a time). If CPUs are idle because coprocs are in use, don't count it towards CPU shortfall.

6) scheduler_cpus() and enforce_schedule() take coprocs and avg_ncpus into account. They try to keep GPUs busy if possible.


This document describes three related parts of the BOINC core client:

CPU scheduling policy
Of the results that are runnable, which ones to execute? BOINC will generally execute NCPUS results at once, where NCPUS is the minimum of the physical number of CPUs (counting hyperthreading) and the user's 'max_cpus' general preference.
CPU scheduling enforcement
When to actually enforce the schedule (i.e. by preempting and starting tasks)? Sometimes it's preferable to delay the preemption of an application until it checkpoints.
Work-fetch policy
When should the core client ask a project for more work, which project should it ask, and how much work should it ask for?

The goals of these policies are (in descending priority):

  1. Results should be completed and reported by their deadline (because results reported after their deadline may not have any value to the project and may not be granted credit).
  2. NCPUS processors should be kept busy.
  3. At any given point, a computer should have enough work so that NCPUS processors will be busy for at least min_queue days (min_queue is a user preference).
  4. Project resource shares should be honored over the long term.
  5. If a computer is attached to multiple projects, execution should rotate among projects on a frequent basis (as defined by the user's 'CPU scheduling period' preference).
  6. Execution should not switch between projects much more frequently than the scheduling period, Otherwise, if the 'remove processes from memory' preference is set, and some applications take a long time to resume from a checkpoint, lot of CPU time will be wasted.

In previous versions of BOINC, the core client attempted to maintain at least one result for each attached project, and would do weighted round-robin CPU scheduling among all projects. In some scenarios (any combination of slow computer, lots of projects, and tight deadlines) a computer could miss the deadlines of all its results. The new policies solve this problem as follows:

  • Work fetch is limited to ensure that deadlines can be met. A computer attached to 10 projects might have work for only a few (perhaps only one) at a given time.
  • If deadlines are threatened, the CPU scheduling policy optimizes the likelihood of meeting deadlines, at the expense of variety.

Concepts and terms

Wall CPU time

Wall CPU time is the amount of wall-clock time a process has been runnable at the OS level. The actual CPU time may be less than this, e.g. if the process does a lot of paging, or if other (non-BOINC) processing jobs run at the same time.

BOINC uses wall CPU time as the measure of CPU resource usage. Wall CPU time is more fair than actual CPU time in the case of paging apps. In addition, the measurement of actual CPU time depends on apps to report it correctly, and they may not do this.

Normalized CPU time

The normalized CPU time of a result is an estimate of the wall time it will take to complete, taking into account

  • the fraction of time BOINC runs ('on-fraction')
  • the fraction of time computation is enabled ('active-fraction')
  • CPU efficiency (the ratio of actual CPU to wall CPU)

but Not taking into account the Resource Share of Projects.

Project-normalized CPU time

The project-normalized CPU time of a result is an estimate of the wall time it will take to complete, taking into account the above factors plus the project's resource share relative to other potentially runnable projects.

The 'work_req' element of a scheduler RPC request is in units of project-normalized CPU time. In deciding how much work to send, the scheduler must take into account the project's resource share fraction, and the host's on-fraction and active-fraction.

For example, suppose a host has 1 GFLOP/sec CPUs, the project's resource share fraction is 0.5, the host's on-fraction is 0.8 and the host's active-fraction is 0.9. Then the expected processing rate per CPU is

(1 GFLOP/sec)*0.5*0.8*0.9 = 0.36 GFLOP/sec

If the host requests 1000 project-normalized CPU seconds of work, the scheduler should send it at least 360 GFLOPs of work.

Result states

R is runnable if

  • Neither R nor R.project is suspended, and
  • R's input files have been downloaded, and
  • R hasn't finished computing

R is nearly runnable if

  • Neither R nor R.project is suspended, and
  • None of R's input files is in a 'download deferred' state.
  • R hasn't finished computing

Project states

P is runnable if

  • P has at least one runnable result (this implies that P is not suspended).

P is downloading if

  • P is not suspended, and
  • P has at least one result whose files are being downloaded and none of the downloads is deferred.

P is fetchable (i.e. the work-fetch policy allows work to be fetched from it) if

  • P is not suspended, and
  • P is not deferred (i.e. its minimum RPC time is in the past), and
  • P's no-new-work flag is not set, and
  • P is not overworked (see definition below), and
  • a fetch of P's master file is not pending

P is latency-limited if

  • The client's last scheduler RPC to P returned a 'no work because of deadlines' flag, and
  • the RPC reply's delay request has not yet elapsed.

This means that P has work available, but didn't send any because the work's deadlines couldn't be met given the existing work queue. P is potentially runnable if

  • P is either runnable, downloading, fetchable, overworked, or latency-limited.

This means that, to the best of the client's knowledge, it could do work for P if it wanted to.

Debt

Intuitively, a project's 'debt' is how much work is owed to it, relative to other projects. BOINC uses two types of debt; each is defined for a set S of projects. In each case, the debt is recalculated periodically as follows:

  • A = the wall CPU time used by projects in S during this period
  • R = sum of resource shares of projects in S
  • For each project P in S:
    • F = P.resource_share / R (i.e., P's fractional resource share)
    • W = A*F (i.e., how much wall CPU time P should have gotten)
    • P.debt += W - P.wall_cpu_time (i.e. what P should have gotten minus what it got).
  • P.debt is normalized so that the mean or minimum is zero.

Short-term debt is used by the CPU scheduler. It is adjusted over the set of runnable projects. It is normalized so that minimum short-term debt is zero, and maximum short-term debt is no greater than 86,400 (i.e. one day).

Long-term debt is used by the work-fetch policy. It is defined for all projects, and adjusted over the set of potentially runnable projects. It is normalized so that average long-term debt, over all project, is zero.

Round-robin simulation

The CPU scheduling and work fetch policies use the results of a simulation of weighted round-robin scheduling applied to the set of nearly runnable results. The simulation takes into account on-fraction and active-fraction. It produces the following outputs:

  • deadline_missed(R): whether result R misses its deadline.
  • deadlines_missed(P): the number of results R of P for which deadline_missed(R).
  • total_shortfall: the additional normalized CPU time needed to keep all CPUs busy for the next min_queue seconds.
  • shortfall(P): the additional normalized CPU time needed for project P to keep it from running out of work in the next min_queue seconds.

In the example below, projects A and B have resource shares 2 and 1 respectively. A has results A1 and A2, and B has result B1. The computer has two CPUs. From time 0 to 4 all three results run with equal weighting. At time 4 result A2 finishes. From time 4 to 8, project A gets only a 0.5 share because it has only one result. At time 8, result A1 finishes.

In this case, shortfall(A) is 4, shortfall(B) is 0, and total_shortfall is 2.

http://boinc.berkeley.edu/rr_sim.png

CPU scheduling policy

The CPU scheduler uses an earliest-deadline-first (EDF) policy for results that are in danger of missing their deadline, and weighted round-robin among other projects if additional CPUs exist. This allows the client to meet deadlines that would otherwise be missed, while honoring resource shares over the long term. The scheduling policy is:

  1. Set the 'anticipated debt' of each project to its short-term debt
  2. Let P be the project with the earliest-deadline runnable result among projects with deadlines_missed(P)>0. Let R be P's earliest-deadline runnable result not scheduled yet. Tiebreaker: least index in result array.
  3. If such an R exists, schedule R, decrement P's anticipated debt, and decrement deadlines_missed(P).
  4. If there are more CPUs, and projects with deadlines_missed(P)>0, go to 1.
  5. If all CPUs are scheduled, stop.
  6. If there is a result R that is currently running, and has been running for less than the CPU scheduling period, schedule R and go to 5.
  7. Find the project P with the greatest anticipated debt, select one of P's runnable results (picking one that is already running, if possible, else the one received first from the project) and schedule that result.
  8. Decrement P's anticipated debt by the 'expected payoff' (the scheduling period divided by NCPUS).
  9. Go to 5.

The CPU scheduler runs when a result is completed, when the end of the user-specified scheduling period is reached, when new results become runnable, or when the user performs a UI interaction (e.g. suspending or resuming a project or result).

CPU schedule enforcement

The CPU scheduler decides what results should run, but it doesn't enforce this decision. This enforcement is done by a separate scheduler enforcement function, which is called by the CPU scheduler at its conclusion. Let X be the set of scheduled results that are not currently running, let Y be the set of running results that are not scheduled, and let T be the time the scheduler last ran. The enforcement policy is as follows:

  1. If deadline_missed(R) for some R in X, then preempt a result in Y, and run R (preempt the result with the least CPU wall time since checkpoint). Repeat as needed.
  2. If there is a result R in Y that checkpointed more recently than T, then preempt R and run a result in X.

Work-fetch policy

A project P is overworked if

  • P.long_term_debt < -sched_period

This condition occurs if P's results run in EDF mode (and in extreme cases, when a project with large negative LTD is detached). The work-fetch policy avoids getting work from overworked projects. This prevents a situation where a project with short deadlines gets more than its share of CPU time.

The work-fetch policy uses the functions

frs(project P)

P's fractional resource share among fetchable projects.

The work-fetch policy function is called every few minutes (or as needed) by the scheduler RPC polling function. It sets the variable P.work_request_size for each project P, which is the number of seconds of work to request if we do a scheduler RPC to P. This is computed as follows:

for each project P
    if P is suspended, deferred, overworked, or no-new-work
        P.work_request_size = 0
    else
        P.work_request_size = shortfall(P)

if total_shortfall > 0
    if P.work_request_size==0 for all P
        for each project P
            if P is suspended, deferred, overworked, or no-new-work
                continue
            P.work_request_size = 1

    if P.work_request_size==0 for all P
        for each project P
            if P is suspended, deferred, or no-new-work
                continue
            P.work_request_size = 1

    if P.work_request_size>0 for some P
        Normalize P.work_request_size so that they sum to total_shortfall
        and are proportional to P.resource_share

For non-CPU-intensive projects, P.work_request_size is set to 1 if P has no nearly-runnable result, otherwise 0.

The scheduler RPC mechanism may select a project to contact because of a user request, an outstanding trickle-up message, or a result that is overdue for reporting. If it does so, it will also request work from that project. Otherwise, the RPC mechanism chooses the project P for which

P.work_request_size>0 and
P.long_term_debt + shortfall(P) is greatest

and requests work from that project. Note: P.work_request_size is in units of normalized CPU time, so the actual work request (which is in units of project-normalized CPU time) is P.work_request_size divided by P's resource share fraction relative to potentially runnable projects.


Scheduler work-send policy

NOTE: the following has not been implemented, and is independent of the above policies.

The scheduler should avoid sending results whose deadlines are likely to be missed, or which are likely to cause existing results to miss their deadlines. This will be accomplished as follows:

  • Scheduler requests includes connection period, list of queued result (with estimated time remaining and deadline) and project resource fractions.
  • The scheduler won't send results whose deadlines are less than now + min_queue.
  • The scheduler does an EDF simulation of the initial workload to determine by how much each result misses its deadline. For each result R being considered for sending, the scheduler does an EDF simulation. If R meets its deadline and no result misses its deadline by more than it did previously, R is sent.
  • If the scheduler has work but doesn't send any because of deadline misses, it returns a 'no work because of deadlines' flag. If the last RPC to a project returned this flag, it is marked as latency-limited and accumulates LTD.

Describing scenarios

We encourage the use of the following notation for describing scheduling scenarios (times are given in hours):

P(C, D, R)

This describes a project with

  • C = CPU time per task
  • D = delay bound
  • R = fractional resource share

A scenario is described by a list of project, plus the following optional parameters:

  • NCPUS: number of CPUS (default 1)
  • min_queue
  • leave_in_memory
  • cpu_scheduling_period

An example scenario description is:

P1(1000, 2000, .5)
P2(1, 10, .5)
NCPUS=4

Scenarios

Scenario 1

P1(0.1, 1, .5)
P2(1, 24, .25)
P3(1, 24, .25)
NCPUS = 2
leave_in_memory = false
cpu_scheduling_period = 1

Typically one CPU will process 6-minute tasks for P1, and the other CPU will alternate between P2 and P3. It's critical that the scheduler run each task of P2 and P3 for the full CPU scheduling period. If we went strictly by debt, we'd end up switching between them every 6 minutes, and both P2 and P3 would have to resume from a checkpoint each time. For some apps (e.g. Einstein@home) resuming from a checking takes several minutes. So we'd end up wasting most of the time on one CPU.