Method and device for optimizing retransmission timeout timer in cluster storage system

Abstract

The invention discloses a method and a device for optimizing a retransmission timeout timer in a cluster storage system. The method comprises the steps of configuring a core clock with the preset precision for the cluster storage system, and acquiring an estimated value of the retransmission timeout timer according to a sampling value of the reciprocation delay time; and adjusting a value of the retransmission timeout timer in the cluster storage system on the basis of the core clock with the preset precision and applying the estimated value of the retransmission timeout timer. According to invention, the cluster storage system is configured with the core clock timer whose precision is at least at a microsecond level so as to set the precision of the retransmission timeout timer to be the microsecond level precision which is the same as that of the reciprocation delay time. Meanwhile, an appropriate estimation method is selected according to the sampling value of the reciprocation delay time and the clustering scale of the cluster storage system to acquire the estimated value of the retransmission timeout timer, thereby selecting an optimized estimated value of the timeout timer to balance influences imposed on the utilization rate of user bandwidth by forged retransmission and transmission timeout.

Description

Technical field [0001] The present invention relates to the technical field of cluster storage, in particular to a method and device for optimizing a retransmission timeout timer in a cluster storage system. Background technique [0002] At present, the data center uses high-speed and low-latency Ethernet to construct a cluster storage system, and provides data access services through TCP / IP network transmission. In the operating cost of a data center, network bandwidth cost is a relatively large cost expenditure. As the number of nodes in storage clusters, distributed file systems, and metadata management clusters organized by high-speed Ethernet continues to increase, the data transmission rate observed on the user side, that is, the network throughput rate, will drop sharply. Called throughput collapse behavior (Incast). [0003] In a cluster storage system, data blocks are stored on multiple different cluster nodes in a striped manner to ensure high availability and high reli...

AI Technical Summary

Problems solved by technology

The research found that the main reason for the sharp decline in the throughput performance of the cluster storage system is the TCP retransmission timeout

[0004] The clock system used by the existing cluster storage system TCP transmission protocol is still a low-precision clock system, which only supports timer calls on the order of milliseconds, resulting in the retransmission timeout timer (RTO) of the cluster storage system TCP transmission protocol min ) can only be kept on the order of milliseconds

When the transient burst packet loss event is not detected through the TCP transmission protocol feedback mechanism, the TCP transmission protocol will decide when to start retransmission according to the value of the timeout timer. The default timeout timer (RTO) in the TCP transmission protocol min ) is generally 200 milliseconds (Linux2.6), and the network round-trip delay (RTT) of the cluster storage system built on the high-speed Ethernet is on the order of microseconds, so a retransmission timeout will make the network link in the " "idle" time is as much as 1000 times the round-trip delay, which not only causes a waste of transmission bandwidth, but also exacerbates the impact of TCP retransmission timeout on throughput collapse behavior

[0005] While the existing pair timeout timer (RTO min ) values ​​are estimated by some methods of estimating the timeout timer (RTO min ) is too low, resulting in a large number of spurious retransmissions, and some estimated timeout timers (RTO min ) is too high, which exacerbates the adverse impact of retransmission timeout on throughput collapse behavior, and cannot find a way to balance the adverse effects of existing estimation methods, so as to select the optimal timeout timer (RTO min ) estimates

Images