1. AN EXTENDED OPENMP TARGETING ON THE HYBRID ARCHITECTURE OF SMP-CLUSTER Author ： Y. Zhao 、 C. Hu 、 S. Wang 、 S. Zhang Source ： Proceedings of the 2nd IASTED international conference on Advances in computer science and technology Speaker ： Cheng-Jung Wu

2. Outline Introduction Extensions in EOMP Computing Resource Definition Hierarchical Data Layout and Data Mapping Execution Model for EOMP Experiments and Results Dot Product Matrix multiplication under EOMP execution model on SMP cluster Conclusion

3. Introduction Clusters of shared-memory multiprocessors (SMPs) More and more popular in High Performance Computing area SMP clusters’ hybrid architectures Supports for a wide range of parallel paradigm Three programming paradigms Standard message passing Hybrid paradigm corresponding to the underlying architecture Shared memory paradigm built on a Software Distribute Software Memory (SDSM) Three major metrics Performance 、 Portability 、 Programmability

4. Introduction None of the three parallel programming paradigms can meet on all of the three metrics New parallel paradigm (EOMP) A compromising model Balance the three major metrics Features Good programmability Acceptable performance Improve memory behavior Data locality The programs running on SMP cluster Inter-node and intra-node data locality

5. Extensions in EOMP Since OpenMP Shared-memory systems Lacks the support for distributed memory system New directives Computing resource definition Data mapping

6. Computing Resource Definition Definitions Virtual node (VN) Virtual processor (VP) VNs Physical nodes Target units of inter-node data distribution VPs Physical processors Target units of intra-node data reallocation and task scheduling during compilation

7. Computing Resource Definition Semantics of computing resource definition directives Examples for processor mapping are given

8. Hierarchical Data Layout and Data Mapping ： Inter-node Data Mapping Scalar data defined in the EOMP Shared data at default Every node gets an own copy of the data Inter-node task parallel Allows the shared scalar data be modified in certain nodes Global addresses of distributed arrays Llocal addresses Inter-node data mapping distributes the mapped arrays to VNs Semantics for inter-node data distribution directive: #pragma eomp distribute a (BLOCK*) onto N

9. Hierarchical Data Layout and Data Mapping ： Intra-node Data Mapping Shared memory data layout takes the advantage of global address Technically No further data mapping is required inside the nodes In certain cases Improper order of data access would decrease cache performance false sharing or long-stride access For instance ： two threads always access the nearby array elements in memory at same time cache performance may be very poor due to severe false sharing Optimizations for intra-node data layout will be necessary

10. Hierarchical Data Layout and Data Mapping ： Intra-node Data Mapping An extreme example Experiment on that circumstance shows an overall 90% reduction of L1 cache miss after the intra-node data reallocation optimization (On 4-cpu IA64 SMP; the array a is of 1M size).

11. Hierarchical Data Layout and Data Mapping ： Intra-node Data Mapping Two strategies can be adopted to reduce cache miss Rearrange the access order of each thread Not always possible for compiler optimization It depends closely on the source program structure In the interleaving data case above, this means to avoid accessing the neighboring data in memory at the same time. Reallocate the data layout in memory, Not change data dependencies of the source program Assures the correctness of this optimization Store the data that accessed by the same thread in a contiguous memory block

12. Hierarchical Data Layout and Data Mapping ： Intra-node Data Mapping Intra-node data reallocation programmer-specified directives compiler reference analysis Intra-node data reallocating data in memory Additional time and space overheads Evaluating the performance speedup of this optimization The data locations have been changed The reallocated data should be forbidden Semantics for intra-node data reallocation directive #pragma eomp distribute a (CYCLIC,*) intra

13. Execution Model for EOMP Inter-node barriers and broadcasts Modifications of shared variables in the parallel section at the edges of task parallel region Maintain data consistency Inter-node communications use explicit message passing

14. Execution Model for EOMP Massage passing & multithreading program generated By compiler first distributes data and schedule the tasks across nodes Then deals with the intra- node data reallocation and task scheduling

15. Experiments and Results ： Dot Product

16. The experiment result shows that the efficiency of the EOMP based on the runtime library is similar to the MPI+OpenMP program (better under some cases) But not good as pure MPI, because the amount of calculations in the dot product operation are not enough, comparing to the cost of inside-node scheduling

17. Matrix multiplication under EOMP execution model on SMP cluster C=A*B A and C is distributed in rows B is distributed in columns

18. Matrix multiplication under EOMP execution model on SMP cluster Matrix size is small The cost of inter-node scheduling and communications are relatively high (compared with the computation cost) The three distributed memory models can not acquire a speedup Matrix size becomes larger The three distributed memory models achieve reasonable speedups

19. Notice that the EOMP model after intra-node data reallocation Gets a high speedup when the matrix size is large Showing that the improved intra-node cache performance can greatly benefit the overall performance of the program on SMP clusters Matrix multiplication under EOMP execution model on SMP cluster

20. Peaks of EOMP-INDR curves in 500*500 and 1000*1000 cases The effect of data reallocation is related with both the size of cache line local b As the nodes become more and more, the size of local b on each node becomes smaller That means the cache line may fill in more rows of local b, thus the cache misses is reduced Explain why the peak in 500*500 multiplication case comes earlier than that of the 1000*1000 case Matrix multiplication under EOMP execution model on SMP cluster

21. Conclusion The experiment result Feasibility of our execution model The benefit gained from intra-node data reallocation For future work, we plan to develop a complete source to source EOMP compiler Be based on ORC (Open Resource Compiler for IA64) Our current runtime library prototype Focusing on the communication generation and data management.