Network switches accessing external memory cloud.

Unlock the Cloud: How Generic External Memory is Revolutionizing Network Switches

Theo Raines in Tech & Innovation August 2025 • 4 min read.

"Discover how a groundbreaking approach to network switch architecture is leveraging remote memory to boost performance and slash costs in data centers."

In today's fast-paced digital world, data centers are the backbone of countless applications, from streaming services to online shopping. These applications demand ever-increasing speeds and efficiency, pushing the limits of existing network infrastructure. Network switches, the unsung heroes of data centers, play a crucial role in directing traffic and ensuring smooth operations. However, traditional switches face a significant challenge: limited memory capacity.

Think of network switches as incredibly fast traffic controllers. They need memory to quickly process and forward data packets. But this memory is expensive and limited, leading to performance bottlenecks. Imagine a highway with too few lanes – traffic jams are inevitable. Similarly, when a network switch runs out of memory, it struggles to handle sudden surges in data, leading to delays and dropped packets. This is where the concept of generic external memory comes in – a game-changing approach that promises to revolutionize network switch architecture.

Imagine expanding the memory of a network switch on demand, drawing from a shared pool of resources. This is the core idea behind generic external memory. By allowing switches to access remote memory, data centers can overcome the limitations of on-chip memory, paving the way for faster, more efficient networks and unlocking a new era of possibilities for data-intensive applications.

The Memory Bottleneck: Why Switches Need More Brainpower

Network switches accessing external memory cloud.

Data center switches have traditionally relied on fast but expensive on-chip memory (SRAM or TCAM) to keep up with demanding network traffic. While this approach offers speed, it severely limits the amount of data the switch can handle at any given time. This limitation becomes a major headache for several key applications:

In-network caches, designed to store frequently accessed data for faster retrieval, suffer from reduced hit rates when memory is scarce. This means more requests have to go to slower sources, slowing down overall performance.

Load balancing: Distributing network traffic evenly across servers becomes difficult, leading to overloaded servers and slower response times.
Network monitoring: Gathering detailed network data for analysis and troubleshooting is hampered, making it harder to identify and resolve performance issues.
Packet Forwarding: Packet forwarding, the core function of a switch, struggles in memory space constraint.

The traditional solution of integrating external DRAM directly into the switch ASIC (Application-Specific Integrated Circuit) has proven costly and inflexible. This approach requires complex wiring and dedicated DRAM modules, adding significant hardware expenses. Moreover, the memory capacity is fixed at the time of manufacturing, limiting scalability. This is like building a house with a pre-determined number of rooms – you can’t easily add more later on.

The Future of Network Memory: Co-design and Open Challenges

The journey towards generic external memory for network switches has just begun, and several exciting challenges and research directions lie ahead. One key area is the co-design of remote memory data structures and the switch data plane. Current commodity switches and RNICs primarily support address-based memory access, limiting the types of data structures that can be efficiently implemented. Future research should explore how to optimize data structures and switch data plane designs to support more complex operations like ternary matching.

About this Article -

This article was crafted using a human-AI hybrid and collaborative approach. AI assisted our team with initial drafting, research insights, identifying key questions, and image generation. Our human editors guided topic selection, defined the angle, structured the content, ensured factual accuracy and relevance, refined the tone, and conducted thorough editing to deliver helpful, high-quality information.See our About page for more information.

This article is based on research published under:

DOI-LINK: 10.1145/3286062.3286063, Alternate LINK

Title: Generic External Memory For Switch Data Planes

Journal: Proceedings of the 17th ACM Workshop on Hot Topics in Networks

Publisher: ACM

Authors: Daehyeok Kim, Yibo Zhu, Changhoon Kim, Jeongkeun Lee, Srinivasan Seshan

Published: 2018-11-15

Everything You Need To Know

How does generic external memory improve network switch performance compared to traditional on-chip memory?

Generic external memory addresses the limitations of on-chip memory like SRAM or TCAM in network switches. Traditional switches rely on these fast but expensive memories, restricting the amount of data they can handle. Generic external memory provides a way to expand switch memory on demand, drawing from a shared pool of resources, leading to faster and more efficient networks. This enhancement directly improves application performance by minimizing delays and dropped packets, especially during data surges.

What happens to key network switch functions like in-network caching and load balancing when memory is limited?

When network switches face memory constraints, several critical functions are impacted. In-network caches suffer from reduced hit rates, forcing requests to go to slower sources. Load balancing becomes difficult, leading to overloaded servers. Network monitoring is hampered, making it harder to identify and resolve performance issues. Packet forwarding, the core function of a switch, also struggles, leading to overall network slowdowns and inefficiencies. Resolving this memory constraint directly improves these functions.

What are the drawbacks of integrating external DRAM directly into a switch ASIC, and how does generic external memory address these?

The traditional solution of integrating external DRAM directly into the switch ASIC has limitations. It's costly due to complex wiring and dedicated DRAM modules. Furthermore, the memory capacity is fixed at manufacturing time, limiting scalability. Generic external memory overcomes these limitations by allowing switches to access a shared pool of memory resources, providing scalability and flexibility without the hardware complexities and fixed capacity constraints of integrated DRAM.

Why is co-design of remote memory data structures and switch data planes important for generic external memory, and what are the implications?

Co-design of remote memory data structures and switch data planes is crucial for maximizing the benefits of generic external memory. Current switches primarily support address-based memory access, limiting the types of data structures that can be efficiently implemented. Optimizing data structures and switch data plane designs to support more complex operations like ternary matching will further enhance the performance and capabilities of network switches using generic external memory. This co-design is an ongoing area of research and development.

Beyond speed, what new network capabilities does generic external memory unlock for data centers and data-intensive applications?

Generic external memory's ability to provide scalable and on-demand memory resources opens doors to new network capabilities. Data-intensive applications benefit from faster and more efficient data processing and forwarding. Enhanced in-network caching improves data retrieval speeds. Improved load balancing optimizes resource utilization across servers. More comprehensive network monitoring provides better insights into network performance and potential issues. These advancements collectively contribute to a more robust and efficient data center environment.