Cache management is critical for Content Delivery Networks (CDNs), impacting their performance and operational costs. Most production CDNs apply static, hand-tuned caching policy parameters at cache servers, such as admission frequency or size thresholds for the Hot Object Caches (HOC) of their system. However, these static policies fall short when a server is faced with unpredictable traffic pattern changes, even when policies employ multiple control parameters/knobs. Recent approaches have proposed learning-based solutions to dynamically adjust policy parameters, but they are limited in action space, caching objectives, or impose high overhead. We propose Darwin, a CDN cache management system that is robust to traffic pattern changes and can flexibly optimize different caching objectives with unrestricted action spaces. Darwin employs a three-stage pipeline involving traffic pattern feature collection, unsupervised clustering for classification, and neural bandit expert selection to choose the optimal caching policy. Through extensive simulations, experiments using an Apache Traffic Server (ATS)-based prototype, and theoretical analysis, we show that Darwin achieves significant performance gain w.r.t. different objectives such as maximizing object hit rates and minimizing disk writes, while simultaneously adapting to traffic pattern shifts. Darwin imposes negligible overhead and achieves high throughput compared to the state-of-the-art.

Ahstract-RDMA over Converged Ethernet (RoCE) has gained significant attraction for datacenter networks due to its compatibility with conventional Ethernet-based fabric. However, the RDMA protocol is efficient only on (nearly) lossless networks, emphasizing the vital role of congestion control on RoCE networks. Unfortunately, the native RoCE congestion control scheme, based on Priority Flow Control (PFC), suffers from many drawbacks such as unfairness, head-of-line-blocking, and deadlock. Therefore, in recent years many schemes have been proposed to provide additional congestion control for RoCE networks to minimize PFC drawbacks. However, these schemes are proposed for general datacenter environments. In contrast to the general datacenters that are built using commodity hardware and run general-purpose workloads, high-performance distributed training platforms deploy high-end accelerators and network components and exclusively run training workloads using collectives (All-Reduce, All-To-All) communication libraries for communication. Furthermore, these platforms usually have a private network, separating their communication traffic from the rest of the datacenter traffic. Scalable topology-aware collective algorithms are inherently designed to avoid incast patterns and balance traffic optimally. These distinct features necessitate revisiting previously proposed congestion control schemes for general-purpose datacenter environments. In this paper, we thoroughly analyze some of the state-of-the-art RoCE congestion control schemes (DCQCN, DCTCP, TIMELY, and HPCC) vs. PFC when running on distributed training platforms. Our results indicate that pre-viously proposed RoCE congestion control schemes have little impact on the end-to-end performance of training workloads, motivating the necessity of designing an optimized, yet low-overhead, congestion control scheme based on the characteristics of distributed training platforms and workloads.