Efficient Messaging for Java Applications Running in Data Centers

Leveraging InfiniBand for Highly Concurrent Messaging in Java Applications

“…The following subsections describe the most important aspects of DXNet and its core which are depicted in Figure 1 and relevant for further sections of this report. A more detailed insight is given in a dedicated paper [8]. The source code is available at Github [1].…”

“…of 7.8 µs, DXNet requires approx. 3.5 µs of the total RTT (the full breakdown is published in our other publication [8]) and the hardware approx. 2.0 µs (assuming avg.…”

“…We do not want the send and receive threads to waste CPU resources and, therewith, decrease the overall node performance. Experiments have shown that simply adding sleep or yield operations highly impacts network latency and throughput and introduces high fluctuations [8].…”

“…This ensures busy looping for the next iterations keeping latency for successive messages and on high loads low. For further details including evaluation results re- fer to our DXNet publication [8].…”

“…We developed the network subsystem DXNet ( §2) which provides transparent and simple to use sending and event based receiving of synchronous and asynchronous messages with transparent serialization of messaging objects [8]. It is optimized for high concurrency on all operations by implementing lock-free synchronization.…”