S3 Object Storage Needs a Tunable Proxy to Overcome Latency in Real-World Networks for Efficient AI Inferencing
San Tomas + Lafayette
Mon Sep 28 | 1:10pm
Abstract
Latency between an S3 client and the storage cluster is a well know limiter of throughput performance. Inserting a tunable proxy between the client and the storage cluster produces what at first blush appears to be paradoxical performance improvements. Adding a proxy is known to increase end to end latency, latency is bad, so why would we want to insert a device that increases latency? in practice, a tunable proxy can actually increase throughput between client and the server by using network tuning techniques to to mask the full end to end latency from the storage cluster allowing it to operate at high levels of performance. This talk will review testing done of multiple vendor's S3 storage systems under latency induced network conditions with and without a tunable proxy.
Related Sessions
Cloud & Object Storage
S3 Object Storage Needs a Tunable Proxy to Overcome Latency in Real-World Networks for Efficient AI Inferencing
Latency between an S3 client and the storage cluster is a well know limiter of throughput performance. Inserting a tunable proxy between the client and the storage cluster produces what at first blush appears to be paradoxical performance improvements. Adding a proxy is known to increase end to end latency, latency is bad, so why would we want to insert a device that increases latency? in practice, a tunable proxy can actually increase throughput between client and the server by using network tuning techniques to to mask the full end to end latency from the storage cluster allowing it to operate at high levels of performance. This talk will review testing done of multiple vendor's S3 storage systems under latency induced network conditions with and without a tunable proxy.
Cloud & Object Storage
Impact of Increasing Drive Density on Magic Pocket
As drive density continues to rise, large-scale storage systems face new tradeoffs that go well beyond simply storing more data per device. Higher-capacity drives can improve space and cost efficiency, but they also increase pressure on rebuild times, recovery behavior, IO bandwidth, background maintenance and overall fleet reliability.
This talk examines how increasing drive density affects Magic Pocket, Dropbox’s custom-built exabyte scale storage infrastructure and how those hardware shifts propagate through the software and operational layers. We will discuss the practical challenges introduced by larger drives, including longer repair windows, changing failure and recovery characteristics and the growing mismatch between per-drive capacity growth and improvements in bandwidth and service time.
We will then show how these constraints influence system design decisions across the stack, including data placement, repair strategy, throughput management and reliability tradeoffs. The goal is to show how evolving hardware characteristics force storage systems to adapt in order to maintain performance, durability and operational efficiency at scale.
By grounding the discussion in lessons from a production storage system, this session will give developers and infrastructure engineers a practical framework for thinking about drive-density growth as a cross-layer design problem spanning hardware, distributed systems and operations.
Cloud & Object Storage
The Next Order of Magnitude: Storage for AI
AI is pushing storage toward a new order of magnitude. Demand is approaching effectively limitless scale, with annual industry opportunity potentially reaching tens of zettabytes per year, while supply, media availability, datacenter growth, and power infrastructure remain constrained. This talk explores what it will take to build storage for the AI era: not only more capacity, but more power, more sites, more mobility of load, and better integration with the communities that host this infrastructure. From bring-your-own-power models to renewed interest in nuclear energy, the path to yotta-scale storage will require the industry to rethink how storage growth, energy systems, and social acceptance fit together.
Cloud & Object Storage
CDMI 3.0: Data Management for the AI Era
CDMI 3.0, the third major revision of the Cloud Data Management Interface, is nearing completion in 2026, with submission to update ISO/IEC 17826 planned for early 2027. This update provides multiple powerful capabilities to enhance data management for AI workloads, including dataset discovery, optimal access method discovery, multiple representation specification, graph relationships for building and traversing storage management knowledge graphs, and support for MCP and skills to allow agents to use CDMI functionality as a standard and open data management protocol.
This presentation will provide an overview of the CDMI 3.0 functionality completed to date, and will include demonstrations of the use cases where CDMI helps solve AI data management challenges, including:
1. As a data management agent, how can I discover AI data repositories?
2. As a data management agent, how can I discover the best way to access AI data?
3. As a data management agent, how can I specify the desired data representation?
4. As a data management agent, how can I traverse and build agent-ready knowledge graphs?
5. As a data management agent, how can I manage storage systems and a standard and interoperable way?
The presentation will conclude with a review of the CDMI extension process and a last-call for contributions to CDMI 3.0.
Cloud & Object Storage
S3-Compatible? Testing the Reality of Object Storage Interoperability
In 2006, AWS introduced the S3 object storage API with little fanfare. 20 years later, S3 has become the de facto standard for object storage and the foundation of a massive Cloud & Object Storage ecosystem. Most, if not all, storage vendors implement S3 as their primary object storage API and yet real-world interoperability issues persist.
This talk will explore why “S3-compatible” does not always mean interoperable, and how those gaps impact developers, applications, and cloud portability. The SNIA Cloud Object Storage Test Tools Technical Work Group will discuss their ongoing efforts to build open-source, vendor-neutral interoperability testing tools that help identify and reduce compatibility issues across S3 implementations. We’ll also cover lessons learned from SNIA’s Cloud Object Storage S3 Plugfests, including how collaborative interoperability testing helps improve consistency across the storage ecosystem.
Attendees will learn:
* How interoperability issues surface in production environments
* What kinds of API inconsistencies are most common
* How open-source testing frameworks can help validate S3 compatibility before deployment
