title |
---|
Development of Scalable Communication Library with Technologies for Memory Saving and Runtime Optimization |
- Pl: Takeshi Nanri (Kyushu University)
- co-Pl: Shinji Sumimoto (Fujitsu Ltd.) and Hidetomo Shibamura (ISIT)
- Research Period: 2011 - 2017
- [Research highlights in pdf]({{ site.baseurl }}/images/crest-nanri.pdf)
- Research website: http://ace-project.kyushu-u.ac.jp/
ACP supports a communication model that transfers data via "Global Memory", that is a virtual memory space managed by the library. Each process can expose a region of its local memory to other processes by registering that region to the global memory.
Copy data from one place in the Global Memory to another. Internally, the library performs appropriate communication operations for each copy operation so that the data is transferred between corresponding registered regions of local memories.
Permutation, increment/decrement and logarithmic operations, on the value stored in the specified place of the Global Memory.
send-receive with 1KB buffer | ACP on InfiniBand | ACP on Tofu/Tofu2 | ACP on Ethernet | |
---|---|---|---|---|
Memory | 10GB | 370MB | 67MB | 34MB |
Explicit creation / free of channels among processes that support one-directional send / receive.
Create / modify / destruct data structures, such as vector, list, deque map and set, on the Global Memory.
Existing MPI programs can call ACP. Also, ACP can be used to connect multiple MPI programs to reduce memory consumption by communicators of MPI
Study behavior of one-sided communications performed with ACP. Packets are generated automatically, and simulated according to the configuration of networks and the pseudo code.