Soft RealTime Scheduling on Simultaneous Multithreaded Processors

1
Soft Real-Time Scheduling on Simultaneous
Multithreaded Processors

• Rohit Jain,
• Christonpher J. Hughes
• Sarita V. Adve
• IEEE REAL-TIME SYSTEMS SYMPOSIUM (RTSS02)
• Present by Yen-Sheng Chang
• Monday, November 24, 2003

2
Outline

• Abstract
• Introduction to SMT (hyperthreading)
• Related Works
• Resource Sharing Algorithms
• Co-Scheduling Algorithms
• Experimental Methodology
• Results
• Conclusions

3
Abstract

• Simultaneous multithreading (SMT) improves
  processor throughput by processing instructions
  from multiple threads each cycle.
• Two Decisions with SMT
• co-schedule selection
• Which threads to run simultaneously (the
  co-schedule)
• resource sharing
• How to share processor resources among
  co-scheduled threads.
• The choice of co-scheduling and resource sharing
  algorithm may be tightly coupled.

4
Abstract (conclude.)

• We find (using simulation) that the best
  algorithm uses global scheduling, exploit
  symbiosis, prioritizes high utilization tasks,
  and uses dynamic resource sharing.
• Significant profiling overhead
• No admission control
• Trade off schedulability!!! (our approach)

5
Introduction to SMT (hyperthreading)
Traditional Pipelining (single-issue)
Superscalar (wide-issue)
2-issue
6
SMT (cont.)
Multithreaded Processor
7
SMT (cont.)
simultaneous multithreaded
traditional (single-issue)
superscalar
multithreaded
8
SMT (cont.)

• Review of two questions
• co-schedule
• resource sharing

• Architecture

9
Related Works

• co-schedule selection
• The fine-grained resource sharing problem is
  unique to SMT.
• ICOUNT (seeks to maximize throughput)
• Without consideration of any real-time deadline.
• Symbiotic Job Scheduling for SMT
• Symbiosis exploited

10
Related Works

• resource sharing
• Transparent Threads
• no real time tasks
• Applications of Thread Prioritization in SMT
• one interactive task with other non-real-time
  tasks.
• Real-Time Scheduling on Multithreaded
  Processors
• no SMT, no resource sharing problems

11
Resource Sharing Algorithms

• The threads share most processor resources
• Instruction fetch mechanism
• Instruction window
• Execution units (functional units)
• Caches
• Previous work has focused mostly on resource
  sharing algorithms to maximize total throughput
  in terms of completed instructions per cycle
  (IPC)
• May have negative/positive impact on the
  schedulability.

12
Resource Sharing Algorithms (cont.)

• Throughput-driven resource sharing
• ICOUNT, gives priority to the thread that has the
  least instructions in the instruction window.
• Refer to Dynamic algorithm.
• Performance prediction is difficult.
• Resource sharing with performance guarantees
• Static resource sharing algorithm where a fixed
  set of resources is reserved for a given job.
• May be suboptimal!
• Performance prediction is easy (identical to
  uniprocessor).
• Resources controlled by thread-specific resource
  sharing algorithms
• SMT is particularly sensitive to the instruction
  fetch bandwidth sharing.
• Fetch bandwidth instruction window.

13
Co-Scheduling Algorithms

• Design Space explored
• Partitioning vs. Global scheduling
• Symbiosis-aware vs. Symbiosis-oblivious
• Prediction of Execution Time, Utilization, and
  Symbiosis
• Partitioning Algorithms
• Global Scheduling Algorhtms

14
Co-Scheduling Algorithms (cont.)

• Partitioning vs. Global scheduling
• Admission control
• Task migration on an SMT processor is free.
• Symbiosis-aware vs. Symbiosis-oblivious
• With dynamic resource sharing on an SMT
  processor, the execution time of a job depends on
  which other jobs are co-scheduled with it.

15
Co-Scheduling Algorithms (cont.)

• Design space explored
• EDF as the underlying algorithm.
• Partitioning ? EDF schedules the tasks within a
  context
• Global scheduling ? EDF chooses the next task.

16
Predicting Execution Time, Utilization, and
Symbiosis

• Review

IPC instruction count
17
Predicting Execution Time, Utilization, and
Symbiosis (cont.)

• IPC
• Job IPC in single-thread mode
• ? profiling one frame of each frame type.
• Job IPC with static resource sharing
• ? profiling each allocation in single-threaded
  mode
• Job IPC with dynamic resource sharing
• ? profiles all possible co-schedules to obtain
  the task IPCs.
• Average job IPC with dynamic resource sharing
• ? when the IPCs depend on the co-schedule, an
  approximation must be made.
• ? we use the job IPC averaged across all
  possible co-schedules are measured above.
• Instruction count
• Use the average instruction count of a large
  number of frames as the prediction.

18
Partitioning Algorithm

• Different between multiprocessor and SMT
• A partition in SMT is a set of tasks such that no
  two will execute simultaneously. (Thus, on an SMT
  supporting N contexts, up to N partitions may be
  created)
• The cost of migrating tasks for SMT is free!
• SMT can allocate resources among the context.

19
Partitioning Algorithm (cont.)

• PART-NOSYM-DYN-b
• Bin-packing based algorithm uses the
  first-fit-decreasing-utilization (FFDU)
  heuristic.
• Approximation for utilization (the need of
  enhanced-version)
• PART-NOSYM-DYN-e
• Modifies the admission test so that no task-set
  is ever rejected in this phase.
• Simulates the schedule for a hyperperiod, to
  determine if it would meet the deadlines.
• Complexity is high, but it gives partitioning the
  fairest showing against global scheduling.

20
Partitioning Algorithm (cont.)

• PART-NOSYM-STAT
• Independent of co-schedule
• Only dependent on resource allocation
• ? No need for enhanced-version!
• FFDU heuristic with an EDF admission test.
• C1, C2, , Cn denote the N hardware contexts.
• Initial all resources are allocated to C1
• Re-allocation resources from C1 to Ck such that
  Ck can accommodate.
• If C1 dont have enough resource ? Fail
• Remove smallest utilization task form C1 to
  another context that can accommodate it.
• If no such context is found ? Fail

21
Partitioning Algorithm (cont.)

• PART-SYM-DYN-b
• Maximizes average symbiosis among tasks in
  different partitions, while keeping the total
  utilization of tasks in each partition reasonably
  balanced.
• Weighted hypergraph with nodes representing the
  tasks
• A hypergraph is a graph in which generalized
  edges (called hyperedges) may connect more than
  two nodes.
• The weight on a hyperedge (u1, u2, , uN) is the
  inverse of the symbiosis factor of the
  co-schedule formed by tasks u1, u2, , uN.
• Each node is weighted with its tasks
  utilization.
• A hypergraph-partitioning algorithm is used.
• The sum of node-weights (utilization) is
  balanced.
• The weight of the hyperedges is minimized
  (maximizing symbiosis)
• PART-SYM-DYN-e

Reference B. L. Chamberlain. Graph Partitioning
Algorithms for Distributing Workloads of
Parallel Computations
22
Global Scheduling Algorithms
23
Global Scheduling Algorithms (cont.)

• Symbiosis-Oblivious Global Scheduling
• GLOB-NOSYM-PLAIN
• EDF
• GLOB-NOSYM-US
• EDF-USm/2m-1 algorithm
• Giving the highest priority to high utilization
  tasks in the task set.

Reference A. Srinivasan and S. Baruah.
Deadline-based Scheduling of Periodic Task
Systems on Multiprcessor
24
Global Scheduling Algorithms (cont.)

• Symbiosis-Aware Global Scheduling
• GLOB-SYM-PLAIN
• Extends EDF to exploit symbiosis in a
  straightforward way.
• It first selects the task with the earliest
  deadline.
• For the other (N-1) tasks, it chooses the set
  that maximizes symbiosis when running with the
  first task.
• Positive ? Improving schedulability (improve
  overall throughput)
• Negative ? Potentially reduce schedulability. (no
  real-time characteristic)
• GLOB-SYM-US
• Improve the negative of GLOB-SYM-PLAIN
• Defaults to GLOB-NOSYM-US if a task Ti has
  utilization Ui gt N/2N-1
• Otherwise, it defaults to GLOB-SYM-PLAIN

25
Co-schedule Algorithm (conclude)
26
Experimental Methodology

• Metrics
• critical serial utilization (CSU)
• The total utilization obtained by uniformly
  increasing the utilization of all tasks until a
  further increase causes the task-set to become
  unschedulable. (5 deadline ? soft real-time)

27
Experiment Setup
RSIM simulator
Real Workload
28
Results

• Best algorithm ? GLOB-SYM-US
• Static vs. Dynamic resource sharing
• Static resource sharing generally implies lower
  throughput than dynamic resource sharing.
• Partitioning vs. Global algorithm
• Enhanced-version is more competitive to
  GLOB-SYM-US.
• Symbiosis-awareness
• Partitioning ? often helps
• Global scheduling ? it depends

29
Conclusions

• Best algorithm
• Global scheduling, exploits symbiosis,
  prioritizes high utilization tasks, uses dynamic
  resource sharing.
• Require a lot of profiling
• Two alternatives
• Partitioning algorithm that utilizes static
  resource sharing
• (PART-NONSYM-STAT)
• Worse schedulability and somewhat more complex.
• Provide a strict admission control and requires
  less profilng.
• Earliest deadline first global algorithms
• (GLOB-NONSYM-PLAIN)
• Not providing strict admission control, but
  requires no profiling.

30
Conclusion (conclude)

• Dynamic resource sharing is better than statis
  for schedulability
• Partitioning algorithm can be made competitive
  with global scheduling algorithm, but with more
  complexity.
• Symbiosis-awareness
• beneficial for partitioning algorithms because
  they do not entirely ignore real-time constraint
• Can hurt or help global scheduling algorithms,
  depending on the relative magnitude of the
  symbiosis factors and total utilization of the
  applications.

31

• Thank you