| 1 |
High-performance networks have developed significantly over the past several |
| 2 |
years. Many of these networks are capable of remote direct memory access |
| 3 |
(RDMA) communication that involves movement of data between processor memories |
| 4 |
without memory copies or remote host processor involvement. |
| 5 |
|
| 6 |
Current high-speed networks deploy powerful communication processor(s) in the |
| 7 |
network interface card (NIC) to handle processing of multiple incoming and |
| 8 |
outgoing messages without interrupting the host processor. For example, the |
| 9 |
Quadrics Elan4 network interconnect can do two overlapping DMAs and allows |
| 10 |
multiple outstanding read/write transactions from and to the network |
| 11 |
interconnect. |
| 12 |
|
| 13 |
Switches used in the modern interconnects also have developed significantly in |
| 14 |
the last several years, facilitating multiple communication paths between |
| 15 |
network endpoints and hence roviding increased levels of concurrency and |
| 16 |
redundancy communication between network endpoints. As a result, the modern |
| 17 |
networks are very capable of handling simultaneous and concurrent data |
| 18 |
movements. |
| 19 |
|
| 20 |
ARMCI allows for its applications to utilize concurrency by its low-overhead |
| 21 |
implementation of one-sided calls |