TASK SCHEDULING ON ADAPTIVE MULTI-CORE
- 1. Task Scheduling on Adaptive Multi-Core Under the guidanc e of KAVITHA n (asst. Professor) Haris M E Roll No. 29 Seminar on :
- 2. contents INTRODUCTION ADAPTIVE MULTI-CORE BACKGROUND WORK BAHURUPI ADAPTIVE MULTI-CORE OPTIMAL SCHEDULE FOR ADAPTIVE MULTI-CORE CONCLUSION FUTURE WORK REFERENCES
- 3. terms • Instruction-level parallelism (ILP) is a measure of how many of the operations in a computer program can be performed simultaneously. • Thread-Level Parallelism (TLP) is a software capability that enables a program to work with multiple threads at the same time.
- 4. INTRODUCTION • Computing systems have made the irreversible transition towards multi-core architectures due to power and thermal limits. • Increasing number of simple cores enable parallel applications benefit from abundant thread-level parallelism (TLP), while sequential fragments suffer from poor exploitation of instruction-level parallelism (ILP) • Adaptive architectures can seamlessly exploit both ILP and TLP.
- 5. Symmetric multi-core • Symmetric multi-core solutions are perfect match for easily parallelizable applications that can exploit significant thread-level parallelism (TLP). • Current trend is to multiply the number of cores on chip to offer more TLP, while reducing the complexity of the core to avoid power and thermal issues.
- 6. Asymmetric multi-core • Asymmetric multi-cores are positioned to accommodate various workload (mix of ILP and TLP) better than symmetric multi- cores. • The asymmetric multi-core architecture provides opportunity for the sequential applications to exploit ILP through a complex core. • As the mix of simple and complex cores is freezed at design time, an asymmetric multi-core lacks the flexibility to adjust itself to dynamic workload.
- 7. Adaptive multi-core • Adaptive multi-core architectures are physically fabricated as a set of simple, identical cores. • At run time, two or more such simple cores can be coalesced together to create a more complex virtual core. • The next step forward to support diverse and dynamic workload is to design a multi-core that can transform itself at runtime according to the applications.
- 8. ADAPTIVE MULTI-CORE… • Adaptive multi-cores appear well poised to support diverse and dynamic workload consisting of a mix of ILP and TLP. • The performance evaluation of the adaptive multi- core only looks at how well a virtual complex core formed with simple physical cores can exploit ILP in a sequential application or TLP in parallel applications, independently.
- 9. Background work The following are few adaptive multi-core architectures have been proposed in the literature recently. • Core Fusion fuses homogeneous cores using complex hardware mechanism. • They evaluate the performance by running serial tasks and parallel tasks separately on the adaptive multi-core.
- 10. Background work…. • Voltron exploits different forms of parallelism by coupling cores that together act as a VLIW processor. • This architecture relies on a very complex compiler that must detect all forms of Parallelism found in the sequential program. • The evaluation is performed with only sequential applications. • Federation shows an alternative solution where two scalar cores are merged into a 2-way out-of-order (ooo) core. • They evaluate the performance by running only sequential applications.
- 11. Bahurupi adaptive multi-core • Bahurupi is a cluster based architecture that imposes realistic constraints on the scheduler. • Figure shows an example of 8-core Bahurupi architecture with two clusters.(C0-C3) and (C4-C7) consisting of 2-way- out-of-order cores. • In coalition mode, the cores can merge together to create a virtual core that can exploit more ILP.
- 12. • A 2-core coalition behaves like a 4-way ooo processor, while a 4-core coalition behaves like an 8- way ooo processor. • Bahurupi can only create coalition within a cluster. • Thus each coalition can consist of at most 4 cores and each cluster can have at most one active coalition at a time. The highlighted cores in Fig. 3 are involved in two coalitions of two and four cores. • Here one parallel application runs its two threads on cores C2 and C3, one medium-ILP sequential application is scheduled to coalition(P0-P1) and one high-ILP sequential application is scheduled to coalition(P4-P7).
- 13. OPTIMAL SCHEDULE FOR ADAPTIVE MULTI-CORE • The ideal adaptive multi-core architecture consist of m physical 2-way superscalar out-of-order cores. Any subset of these cores can be combined together to form one or more complex out-of-order cores. • If r cores(r<=m) form a coalition, then the resulting complex core supports 2r-way out-of-order executions. • The architecture can support any no of coalitions as long as the total number cores included in the coalition at any point does not exceed m.
- 14. The scheduling problem: • Consider an ideal adaptive multi-core architecture consisting of m homogeneous independent physical processors {P0, P1, P2,…,Pm-1} running a set of n (n< = m) preemptive malleable tasks {T0, T1,T2,…,Tn-1}. We assume that the core coalition does not incur any performance overhead. The objective is to allocate and schedule the tasks on the core so as to minimize the makespan Cmax =maxj{Cj} where Cj is the finish time of task Tj. • The adaptive architecture allows both sequential and parallel applications to use time varying number of cores.
- 15. • Let us denote the number of processors assigned to the task Tj by rj, where 0<rj<=m. • Then the set of feasible resource allocations of processors to task is a follows R = {r = (r0,…,rn-1 ) | rj>0, Σrj<=m} n-1 j=0
- 16. Task scheduling in bahurupi Task Scheduling constraints: When scheduling tasks on a realistic adaptive multi-core architecture, we must take into consideration all the constraints and limitations imposed by the system. • Constraint 1(C1): A sequential application can use at most four cores.
- 17. • Constraint 2(C2): It can form at most m/4 coalitions at any time. • Constraint 3(C3): A sequential application can only use cores that belong to the same cluster. • There is no such constraints for the parallel tasks. A parallel task may use any number of available cores to minimize the overall timespan.
- 20. The experiments • They conducted a study to characterize the performance limit of adaptive multi-cores compared to static multi-cores. • They generated different workload (task sets) consisting of varying mix of ILP and TLP tasks. • Used 850 task sets and computed the makespan of each task set on different multi-core architectures.
- 21. • They modeled seven different static and adaptive multi- core configurations in the study as shown in Table.
- 22. Results
- 23. • The normalized speedup of adaptive architectures ranges from10% to 49%. • On an average, adaptive architectures outperform the asymmetric configurations A1, A2 and A3 by 18%, 35% and 52% respectively. • When compared with the symmetric configuration S2, the adaptive architecture performs 26% better, which makes the symmetric configuration S2 a better option than the asymmetric configurations A2 andA3.
- 24. • The results clearly demonstrate that adaptive architectures (Ideal and Bahurupi) perform significantly better when compared to static symmetric and asymmetric architectures. • The performance of Bahurupi is practically identical to that of Ideal adaptive architecture even though Bahurupi imposes certain constraints on core coalition.
- 25. limitations • Adaptive multicore architectures require complex scheduling algorithms unlike static multicores. • Assumption that the core coalition does not incur any performance overhead is not true. • In bahurupi adaptive multi-core architecture core coalitions are limited by realistic constraints which makes it perform lesser than ideal adaptive multi- core architectures.
- 26. conclusion • They have presented a comprehensive quantitative approach to establish the performance potential of adaptive multi-core architectures compared to static symmetric and asymmetric multi-cores. • This is the first performance study that considers a mix of sequential and parallel workloads to observe the capability of adaptive multi-cores in exploiting both ILP and TLP.
- 27. • They employed an optimal algorithm that allocates and schedules the tasks on varying number of cores so as to minimize the timespan. • The experiments reveal that both the ideal and the realistic adaptive architecture provide significant reduction in timespan for mixed workload compared to static symmetric and asymmetric architectures.
- 28. FUTURE WORK • Developing optimal scheduler for higher versions of adaptive multicore architectures. • Improving the scheduler algorithm for better utilization of processors. • Improving bahurupi adaptive multicore architecture more close to ideal adaptive multicore architecture.
- 29. REFERENCES • P.-E. Bernard, T. Gautier, and D. Trystram,“Large Scale Simulation of Parallel Molecular Dynamics,” Proc. 13th Int’l Symp. Parallel Processing 10thSymp. Parallel Distributed Processing, pp. 638-644, 1999. • C. Bienia, S. Kumar, J.P. Singh, and K. Li,“The PARSEC Benchmark Suite: Characterization and Architectural Implications,”Proc. 17th Int’l Conf. Parallel Architectures Compilation Techniques, pp. 72-81, 2008. • E. Blayo and L. Debreu, “Adaptive Mesh Refinement for Finite- Difference Ocean Models: First Experiments,”J. Physical Oceanogra- phy, vol. 29, pp. 1239-1250, 1999.