Figure 4 shows that POP performed up to 1.72% faster in our 236-node InfiniBand cluster when the tickless mode was turned on. The fact that the performance gains increased with the increasing scale suggests that a tickless Linux kernel would mitigate the scalability issues related to OS noise at extreme scales as described in the literature. We also found that application runtime variability was nearly diminished with the tickless kernel, which we consider invaluable.
As we mentioned above, the downside to removing the ticks is the unprocessed bottom half handlers. Consequently, our tickless kernel did not allow the applications to make progress when we used the Ethernet network in the cluster. We then modified the kernel to process all network packets in the OS cores, however, this change reduced the network traffic bandwidth such that applications were slowed down by about 10% when the no tick mode was on. Our future work includes optimizing this behavior with increased batching on the OS cores as we described above.