Presentation on theme: "D YNAMIC T HREAD A SSIGNMENT ON H ETEROGENEOUS M ULTIPROCESSOR A RCHITECTURE Hüsnü Şensoy."— Presentation transcript:
D YNAMIC T HREAD A SSIGNMENT ON H ETEROGENEOUS M ULTIPROCESSOR A RCHITECTURE Hüsnü Şensoy
A GENDA Problem Definition Benchmarking Over EV 5 & EV 6 Architectures Assignment Policies Simulation Results Questions & Comments
P ROBLEM D EFINITION Dynamic Thread Assignment on Heterogeneous Multiprocessor Architecture
P ROBLEM D EFINITION (S TATIC & H OMOGENOUS ) P1 P2 P3 P4 T1 T2 T3 T4 T5 T6 Which thread should be scheduled over which processor ?
P ROBLEM D EFINITION (S TATIC & H ETEROGENOUS ) P1 P3 P2 P4 T1 T2 T3 T4 T5 T6 Which thread should be scheduled over which processor ?
P ROBLEM D EFINITION (D YNAMIC & H ETEROGENOUS ) P1 P3 P2 P4 T1 T2 T3 T4 T5 T6 Which thread should be scheduled over which processors ?
P ROBLEMS How to decide which thread should be scheduled on which processor in an heterogeneous processor environment ? Are heterogeneous architectures better than homogenous ones ? Is dynamic allocation better than static one ?
W ORKLOAD D EFINITION ProgramDescription gzipData compression utility gccC compiler craftyChess program parserNatural language processing bzip2Data compression utility wupwisQuantum chromodynamics swimShallow water modeling mgridMulti-grid solver in 3D potential field galgelFluid dynamics analysis of oscillatory instability equakeFinite element simulation: earthquake modeling lucasNumber theory: primality testing
S IMULATION R ESULTS Dynamic Thread Assignment on Heterogeneous Multiprocessor Architecture
S IMULATION R ESULTS (S TATIC S CHEDULING ) Static Random Assignment Random vs Pseudo Optimal Assignment
S IMULATION R ESULTS (R OUND R OBIN D YNAMIC A SSIGNMENT )
S IMULATION R ESULTS (IPC D RIVEN D YNAMIC A SSIGNMENT )
C OMPONENTS O F S PEED - UP (IPC D RIVEN D YNAMIC A SSIGNMENT )
C OMPONENTS O F S PEED - UP (R OUND R OBIN D YNAMIC A SSIGNMENT )
C ONCLUSION Heterogeneous multiprocessor systems allows to Efficiently accommodate different degrees of thread parallelism Meet the needs of multi-programmed computing environments. The presence of many low area cores ensures a high level of parallelism. The presence of a few high performance cores guarantees high throughput when thread parallelism is low. In this work it is also argued that the benefits of heterogeneous CMP are increased by the use of dynamic policies for assigning threads to processors. This allows to capture the dynamic behavior of the running threads. maximize the usage of the high performance cores. According to the results of the work, a dynamic policy on a heterogeneous CMP can outperform a random assignment policy by 20% to 40%. a homogeneous configuration by 20% to 80% depending on the number of threads simulated.
Q UESTIONS & C OMMENTS Dynamic Thread Assignment on Heterogeneous Multiprocessor Architecture ?
Your consent to our cookies if you continue to use this website.