SNIA Developer Conference September 15-17, 2025 | Santa Clara, CA
Exploring Optimization of Content Delivery Networks
Abstract
Building Content Delivery Networks, or CDNs, requires distributed, localized collections of compute, memory and storage. CDNs are built out of groups of servers at a variety of locations with various tiers and types of caches. Modern CDN caches present a huge array of variable configurations.
Magnition sponsored a university research project to study configurations and algorithms that are used in CDNs to optimize response speed and cost of the front-end caches. By testing the configurations and algorithms through simulations, over 100,000 variants were examined and compared using traffic traces provided by genuine CDN companies. These simulations were able to test what would have been years worth of traffic in the course of roughly a week.
This session discusses the process of simulating specific CDN configurations. We demonstrate how modular components and algorithms are sampled, share some of the results we found, and talk about some of the anomalies we stumbled across along the way. We also demonstrate how to use real-world CDN traces to build realistic scenarios and show graphical representations of the results.
You won't want to miss this session.
Learning Objectives
Content Delivery Networks are intricate and difficult to tune accurately.
Simulations allow a vast experimental space to be examined in reasonable time.
There is quite a bit of room to improve on CDN performance and cost from baseline production setups.
Related Sessions
Distributed Data Placement for NVMf Scale-out Storage Cluster on AWS
Simplyblock builds a distributed lock storage system, which is fully containerized, runs on most virtualization platforms and clouds and is currently optimized for aws environments. Simplyblock is built to fully utilize local nvme IOPS and provide low storage access latency (less than 200 microseconds over the network). Virtual volumes are highly available and protected by erasure coding. Simplyblock clusters are self-balancing and highly scaleable. For this purpose, we use a distributed data placement algorithm. In this session, we plan to discuss the common challenges of a distributed architecture for reliable, self-balancing storage scale-out, specific challenges in the context of low latency and high IOPS density, nvmf-tcp and nvmf multipathing, and present the design of new algorithm to overcome them.
Exploring Optimization of Content Delivery Networks
Building Content Delivery Networks, or CDNs, requires distributed, localized collections of compute, memory and storage. CDNs are built out of groups of servers at a variety of locations with various tiers and types of caches. Modern CDN caches present a huge array of variable configurations.
Magnition sponsored a university research project to study configurations and algorithms that are used in CDNs to optimize response speed and cost of the front-end caches. By testing the configurations and algorithms through simulations, over 100,000 variants were examined and compared using traffic traces provided by genuine CDN companies. These simulations were able to test what would have been years worth of traffic in the course of roughly a week.
This session discusses the process of simulating specific CDN configurations. We demonstrate how modular components and algorithms are sampled, share some of the results we found, and talk about some of the anomalies we stumbled across along the way. We also demonstrate how to use real-world CDN traces to build realistic scenarios and show graphical representations of the results.
You won't want to miss this session.
Problems with AI Interconnects? Come yell at our Opinionated Experts!
AI is driving the need for more data movement between devices, boxes, and racks. This data transport requires higher bandwidths which puts strain on the interconnects as well as host and device designs. While you could tackle these problems on your own, instead come and ask questions of our expert panel from the SFF Technical Work Group to see if they can address your concerns. This will be an open Q&A where we want to hear your concerns so we can address them through industry aligned solutions.
The Promise of NVMe Flexible Data Placement in Data Center Sustainability
By 2040, data centers could account for up to ~14% of all carbon emissions worldwide. Flash SSDs contribute to data center carbon emissions due to their limited endurance. The common practice of over-provisioning Flash SSDs to control write amplification, improve lifetime and optimize for total cost of ownership of data centers is inefficient and contributes to the carbon footprint. Considering the increased focus on data center climate impact and net-zero carbon goals, improving Flash SSD lifetime and utilization are key to reducing carbon emissions. Targeted data placement on SSDs is known to reduce write amplification and improve lifetime. The latest data placement technology - NVMe Flexible Data Placement (FDP), shows great promise in reducing carbon emissions with minimal engineering effort. Providing control to applications to place their data with NVMe FDP results in more informed data placement strategies with minimal modifications to the application stack. Consequently, this leads to reduced SSD write amplification and enables better utilization. In this talk, we explore NVMe FDP’s future in Data Center Sustainability by showcasing its ability to reduce embodied (endurance) and operational (power) carbon emissions for systems at scale. We use several example workloads and CacheLib - a deployed system at scale, to illustrate the impact and reduction in carbon emissions with NVMe FDP.
Nuances in FDP Implementation
With Flexible Data Placement (FDP) from NVM Express® (NVMe) finalized and increasing in ecosystem momentum, it has become clear that implementation choices are becoming a real differentiator among FDP drive configurations. Some customers leverage years of Multi-Streams deployment to migrate onto large RU sizes within a single RG. Some customers may be eager to mirror their experience with Open Channel SSDs by requesting small RGs and RU sizes. Yet another customer base might be examining FDP from a history of working with high Zone counts translating into large RUH counts. This presentation will provide customers with some high level guidance and background while engaging with SSD vendors. By understanding the drive impacts of such FDP configuration choices, customers and vendors can arrive at the best system solution for varying use-cases.
No More LRU: Simple Scalable Caching with Only FIFO Queues
- Caching is used in almost every component of today's storage systems to speed up data access and reduce data movement. The most important component of a cache is the eviction algorithm that decides which objects to keep in the very limited cache space. While Least-Recently-Used (LRU) is the most common eviction algorithm, it suffers from many problems.
- First, LRU is not scalable. LRU maintains objects in last-access order, which requires a doubly-linked list. Each data read request requires moving the requested object to the head of the linked list under locking. Therefore, LRU's throughput cannot scale with the number of cores and cannot benefit from modern multi-core CPUs.
- Second, existing state-of-the-art eviction algorithms are mostly built upon LRU, using one or multiple LRUs with complicated adaptive algorithms. As a result, they are not scalable and are often complex and hard to debug.
- As a cache eviction algorithm, FIFO has many attractive properties, such as simplicity, speed, scalability, and flash-friendliness. However, its most prominent criticism is its low efficiency (high miss ratio).
- In this talk, I will introduce a simple, scalable FIFO-based algorithm with three static queues (S3-FIFO). Evaluated on 6594 cache traces from 14 datasets, I will show that S3-FIFO has lower miss ratios than state-of-the-art algorithms across traces. Moreover, S3-FIFO's efficiency is robust --- it has the lowest mean miss ratio on 10 of the 14 datasets. FIFO queues enable S3-FIFO to achieve good scalability with 6× higher throughput compared to optimized LRU at 16 threads.
- The insight is that most objects in today's storage workloads will only be accessed once in a short window, so it is critical to evict them early (also called quick demotion). The key of S3-FIFO is a small FIFO queue that filters out most objects from entering the main cache, which provides a guaranteed demotion speed and high demotion precision.
