Please note: This PhD defence will take place in DC 2314.
Siyao Zhao, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Ali Mashtizadeh
Modern large scale services distribute each user request across deep service trees, making end user latency sensitive to the tail latency of individual nodes. Reducing this latency motivates revisiting kernel event facilities that predate modern hardware. Kernel bypass offers an alternative but relies on polling due to limitations in processors’ interrupt delivery. This motivates improving processor extensibility and enabling rapid innovation of processor extensions through software rather than hardware.
Meanwhile, the scale of these services demands better performance and density per machine, motivating multisocket systems and a reassessment of conventional assumptions about non-uniform memory access (NUMA) overheads on modern hardware. This thesis addresses these challenges across the system stack, from kernel event handling and processor extensibility to NUMA resource utilization.
In the kernel layer, Schedulable Kqueue (SKQ) is a new scalable design to FreeBSD Kqueue that improves application tail latency and low latency throughput via event scheduling. SKQ offers multiple scheduling policies that improve cache locality and reduce workload imbalance.
For processor extensibility, Cobalt enables developers to access low level architectural features with microcode level overhead and implement efficient architectural extensions in high level languages such as C.
We also quantify the effects of NUMA on IO heavy workloads to improve resource utilization on multisocket systems by running our benchmarking system on various architectures. To our surprise, the results indicate that NUMA does not significantly affect network IO latency and throughput on processors with modern and mature interconnects.