Spin lock killed the performance star

• Published in: 2015 34th International Conference of the Chilean Computer Science Society (SCCC) pp. 1 - 6
• Main Authors: Rivera, Diego, Blasco, Sebastian, Bustos-Jimenez, Javier, Simmonds, Jocelyn
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-11-2015
• Subjects: Internet; Kernel; Linux; Message systems; Multicore processing; Servers; Sockets
• DOI: 10.1109/SCCC.2015.7416588

The Internet has grown quite quickly, requiring more and more processing power each year to handle user requests in a timely fashion. In the multicore world, the addition of server-side threads should help improve server performance. However, several studies have shown that this is not true, identifying the Linux kernel as the possible culprit. Our working hypothesis is that the kernel does not provide a scalable interface for network communications. Through various tests, we narrowed the problem down to the implementation of the spin lock mechanism (a synchronization structure used mostly at the kernel level), which has been inherited from early versions of the Linux kernel. It is only now, with the emergence of multicore architectures, that users have begun to notice the performance hit that the existing spin lock implementation has on parallel systems, especially in multithreaded network protocols. Thus, our recommendation is that spin locks be redesigned so that the full power of multicore systems can be harnessed.