Presentation Information
[A-10-09]Evaluation of CPU Architecture Impact on ECS Fargate Cold Start Latency
◎△Takao Tokunaga1, Atuya Kanno2, Kaichi Taguchi2, Nagisa Nasu2, Akihito Kohiga1, Takahiro Koita2 (1. Doshisha Univ., 2. Graduate School of Doshisha Univ.)
Keywords:
Amazon Web Service,ECS,cold start
In ECS Fargate's on-demand launch mode, a new container task starts for each request, incurring cold start latency at startup. This study empirically evaluated, for both the amd64 and arm64 architectures supported by ECS Fargate, how much the cold start latency differs between them and which phases are affected by the CPU architecture. We defined cold start latency as the time from sending an SQS message to the start of the in-container process, recorded via CloudWatch Metrics, and measured each phase—host provisioning, image pull, and container startup—using the ECS DescribeTasks API. With two factors, architecture (amd64, arm64) and image size (six levels, 100 MB–2 GB), we measured latency 100 times for each of the 12 conditions. Under all conditions, arm64 showed shorter cold start latency than amd64. Host provisioning took about 11 seconds and was nearly constant across architectures, while image pull time grew with image size. The reduction by arm64 ranged from 6.1% (100 MB) to 3.4% (2 GB), and the absolute difference widened with image size from about 1.1 to 2.3 seconds.