Mitigation of Cold Start Delays in Serverless Computing Through Adaptive Heuristic-Based Prewarming Framework

Abstract

Scalable and event-driven execution is offered in serverless computing by abstracting infrastructure management, allowing applications to be deployed and operational control to be easily achieved. There is however the cold start phenomenon which still undermines performance especially on latency-sensitive applications. After a period of idleness of a function, the platform unloads its execution environment, which has led to a higher response time when the same function is invoked again. The current idle-based prewarming solutions maintain functions by activating them periodically, which have minimized cold starts but have also created the problem of excessive resource usage and unnecessary execution expenses. Solutions that are based on learning enhance flexibility but introduce computational overhead, which is incompatible with lightweight serverless platforms. Existing cold start mitigation strategies do not have a balanced mechanism that responds to the real time invocation behavior and minimizes overhead. Majority of the methods consider all functions equally, resulting into over-prewarming or retarding implementation. An adaptive strategy with low overheads that distinguishes functions by frequency of use and selectively implements prewarming within memory and cost limits is still needed. The proposed work suggests an Adaptive Heuristic-Based Prewarming model that classifies dynamically serverless functions into hot, warm and cold ones based on real-time invocation frequency and past access patterns. Hot functions are prewarmed in advance so that they can be executed immediately. Selective prewarming of warm functions is only used when the cold start probability is larger than a predefined threshold but the cold functions are reloaded on demand through the use of checkpoint-based loading. A caching system based on the LRU procedure will store the most frequently used instances and will evict the least used instances when the memory pressures occur. The framework is model-free and therefore ensures real-time flexibility with low computational requirements. Experimental analysis with 10,000 real-world logs of function invocation under cold start conditions demonstrates that the suggested Adaptive Heuristic-Based Prewarming method decreases cold start by 72, which is by way of comparison lower than cold start decreased by existing learning and cache-based techniques (48-63). The mean reaction time is reduced by half to 140ms and this is a 70 % improvement. The cache hit rate is 85 and it has improved by a margin of about 4-9 than the best baseline. Unneeded prewarming overhead is cut down to 18.2 which translates into an approximate 65% cut in execution cost. These findings verify that the proposed approach provides a successful trade-off between reduction in latency, the use of resources and cost efficiency in the case of dynamically serverless workloads.