Influence of the direct message search mechanism based on the TCP protocols on the exchange process

Abstract

The direct search mechanism is implemented with the expansion of the traditional socket TCP interface for receiving messages while bypassing the traditional order of the established queue. This mechanism can be used for high-performance and clustered computer systems in order to intensify data exchange and continuous support of a maximum load on computing machines. The interface for direct message search is implemented on the base of the Linux kernel. The experimental test results are obtained by using a set of simple microbenchmarks. During the test, the sender sends the required number of fixed-size messages via an established connection, and the receiver skips unexpected messages and reads the expected one into the user's space. The approach to finding the expected messages is realized with multiple searches for a case where the socket application treats the TCP socket as a list of messages with the ability to receive and delete the data not only from the buffer top but from any place in it. All expected messages are recognized and processed by the developed seek_recv() call. Each test contains ~80 repetitions, which include such operations as socket opening, sending 800–1000 messages according to the acceptance politics, and socket closure. The system only uses one active socket at the same time. The received results confirm a noticeable decrease in message processing CPU time by 36–40 % and overall productivity growth. However, when the volume of messages is approaching 1000 bytes, i.e. close to the typical size of the TCP packet useful load, there is a productivity decrease in the message exchange process.