Perception-based Congestion Control Mechanism For Distributed File System

In the distribute file system, one data server needs to handle I/O access requests fromplenty of clients at the same time. If the data server does not have enough resources tohandle all the requests arrived and requests that are about to reach, a large number of I/Oaccess requests will be suspended in the data server side, giving rise to congestion.Furthermore, this will lead to severe data server performance degradation. Meanwhile, thedegree of congestion would be enlarged by retransmission over timeout, resulting in dataserver’s throughput decreasing to a small part of its original processing capability. Itbecomes important that how to handle the congestion due to overload and avoid the rapiddecrease of system’s performance, availability and throughput.Taking account the characteristic of distributed file system, this paper proposes acongestion control scheme PbCCM(Perception-based Congestion Control Mechanism).Firstly a high concurrent processing model is designed and implemented which is helpfulto improve service capability of the system, delaying the critical point of the congestion.Then a kind of feedback mechanism between the client and server is introduced, the servercan notify its status to the client when it is overloaded. Then the client will takeappropriate measures to adapt to the data server’s load, gradually reducing the load on thedata server. Finally, additional overhead with the retransmission can be cut down byoptimizing retransmitted write requests.Test results show that, compared with no congestion control scheme of distributedfile system, in the situation of data server congestion, PbCCM can effectively guaranteethe system service performance, mainly reflected in reducing7%~20%of the I/O requestaverage response time and improving25%of the system I/O throughput rate, and greatlydecreasing relatively30%of the request timeout retransmission rate in the situation whenI/O requests can’t get timely response.