Skip to content

ISCA

2023

MXFaaS: Resource Sharing in Serverless Environments for Parallelism and Efficiency

Serverless computing, cloud computing, resource management

Real-world serverless workloads analysis: Azure's serverless functions and Alibaba's microservices.

Open-source serverless benchmarks: FunctionBench, TrainTicket Application and SeBS.

2 inefficient patterns:

  • synchronous I/O within a function
  • functions calling functions

4 workload and execution characteristics:

  • invocations of the same function are bursty
  • idle time dominates function execution
  • substantial state replication in memory
  • concurrent accesses to the same storage locations

solution:

  • Fork handlers and make OS only schedule the oldest N ready-to-execute invocations of the function. Requests that are (1) blocked on I/O and (2) lack a core to run are be buffered.
  • dispatcher combines remote storage accesses and remote function calls from multiple handlers running invocations of the same function.
  • merge storage access is much like a cache access (forward read)
  • dispatcher first initializes the MXContainer state and after that, on demand, spawns a handler process per function invocation.

2016

ARM Virtualization: Performance and Architectural Implications

computer architecture; hypervisors; operating systems; virtualization; multi-core; performance; ARM; x86

Measures the performance overhead of ARM virtualization extensions (VE). And extract the comparison between x86 and ARM hypervisors.