We achieved the second position and the first position in the world rankings of the Graph500 Benchmark and the Green Graph500 Benchmark.
The nineth Graph500 list and the fourth Green Graph500 list were announced at the Supercomputing Conference 2014 (SC14) in New orleans, US. We won the second position and the first position in the world rankings of the Graph500 Benchmark and the Green Graph500 Benchmark. These results show that our algorithm and implementation is efficient and scalable for big data computation in several computing environments, such as distributed-memory supercomputers, shared-memory supercomputers, Intel/AMD servers, and Android devices.
◯ Graph500 List
We have achieved 19585.2 GTEPS for Kronecker graph with SCALE 40 on 82,944 (of 88,128) nodes of the K computer by our MPI-based implementation. It was obtained by our new implementation, which improved a limitation as the number of nodes must be a power of two in old code. However, the Sequoia improves 23,751 GTEPS for SCALE 41 on 82,944 nodes and 663,552 cores. Therefore, the rank of K is the second position. And we also updated the TEPS scores on distributed-memory supercomputers TSUBAME 2.5, as 1344.95 GTEPS for SCALE 36. On the other side, on shared-memory computers, our OpenMP-based implementation achieves 174.704 GTEPS on two racks of UV 2000. This score indicates a fastest of single-node system and a highest TEPS per core in an upper position than this. Futhermore, this implementation also obtains 55.741 GTEPS and the fastest single server for SCALE 32 on Huawei RH5885H V3 with 4-way 15-core Xeon and 2 TB RAM.
|2||K computer||RIKEN Advanced Institute for Computational Science (AICS)||82,944||663,552||40||19585.2|
|12||TSUBAME 2.5||GSIC, Tokyo Institute of Technology||1,024||12,288||36||1344.95|
|42||ismuv2k2 (SGI UV 2000)||The Institute of Statistical Mathematics||1||1,280||33||174.704|
|42||Huawei RH5885H V3||Kyushu University||1||60||32||55.741|
◯ Green Graph500 list
＊ Big Data Category (more than or equal to SCALE 30)We defended the first position of the big data category in the fourth Green Graph500 list. The #2 to the #5 entries in this list are our results. The entries #1, #2, and #3 ware obtained on large memory machine by our OpenMP-based implementation. And we updated the first position entry as 61.48 GTEPS. In particular, the third entry is reasonable for the TEPS (55.7 GTEPS) and TEPS/W (44.42 MTEPS/W) scores on a large problem (SCALE 32). Furthermore, the entries #4 and #5 are obtained by a single node with off-loading techniques using NVM (Non-valatile memory) extension, these can handle large problems without supercomputers.
|4||35.87||Tokyo Institute of Technology||GraphCREST-Custom #1||32||10.6|
|5||28.88||Tokyo Institute of Technology||MEM-CREST Node #2||30||8.0|
The certification in Big data category (Nov. 19, 2014)
＊ Small Data Category (less than SCALE 30)In this category, we have also many entries on various computers such as Intel/AMD servers and Android devices Xperia SO-01F. The entry #2 achieves a power-efficient performance as 235.15 MTEPS/W. These results represent a potential that our graph processing techniques can exhibit a high performance even in the next generation supercomputer, which has power-saving processors.
|4||204.38||Tokyo Institute of Technology||EBD-GoldenBox-Prototype||21||1.64|
◯ Graph500 and Green Graph500 Collaborative Research TeamKatsuki Fujisawa (Research Director) and Yuichiro Yasui from Institute of Mathematics for Industry, Kyushu University, Toyotaro Suzumura form University College Dublin, Hitoshi Sato, Koji Ueno, Keita Iwabuchi and Ryu Mizote from Global Scientific Information and Computing Center, Tokyo Institute of Technology.
Photo in SC14, New Orleans, US (Nov. 19, 2014)