MLED Team

Prateek Jain

Boston University

Arash Sarabi

Arizona State University

Abraham Matta

Boston University

Violet R. Syrotiuk

Arizona State University

Overview

The Multi-Level Error Detection framework, denoted MLED(n, P), is defined by n ≥ 3 levels and a set of policies P. Each level consists of one or more layers, and each layer is governed by a configurable policy over its scope. The architecture generalizes the traditional two-layer approach to error detection by introducing additional levels that can detect and localize errors before corrupted data reaches the final destination.

MLED is designed to be modular and decoupled. In principle, different layers can be configured with their own error-detection, routing, addressing, flow-control, and recovery behavior. The current implementation focuses on error detection for large-scale data transfers, but the architecture is intended to support broader communication functions in the future.

Why MLED Is Needed

• Packet sizes have grown substantially, increasing the chance that errors may slip past conventional transport- and link-layer checks.
• Large file transfers are especially vulnerable, since a single undetected error can invalidate a petabyte-scale scientific dataset and force expensive retransmission.
• Traditional end-to-end recovery is inefficient when corruption is detected only after the full transfer reaches the destination.
• In-network detection enables localization, allowing recovery at intermediate scopes rather than retransmitting the full file from the source.

Key Features

Benefits

• Reduced undetected error probability: adding extra levels reduces the effective undetected error probability across the transfer path.
• Efficient recovery: retransmissions can be triggered from intermediate scopes instead of always restarting from the source.
• Configurability: different policies and payload sizes can be selected to satisfy a target admissible undetected error probability.
• Strong experimental validation: the architecture has been implemented and evaluated on the FABRIC testbed.

MLED Architecture

In MLED, each layer starts and ends with its own source and destination processes, while relay processes connect lower layers to realize higher-level logical communication. This recursive organization allows routing and recovery to be handled locally at each level while still achieving an end-to-end transfer objective.

MLED(4, P) configured across five nodes with recursive levels and relay processes.

Implementation Highlights

Experimental Validation on FABRIC

The framework was evaluated on the FABRIC testbed using a five-node deployment spanning sites in Chicago, New York, Washington, Atlanta, and Dallas. The implementation was compared against the traditional two-layer approach under an adversarial error model that injects errors able to evade both CRC and Internet checksum checks in the baseline design.

• In the traditional approach, such adversarial corruption led to a corrupt file transfer and full retransmission from the source.
• With one additional MLED level using CRC-8, the framework detected and corrected these errors inside the network.
• Under non-zero error rates, MLED achieved a 100% gain in goodput over the traditional approach.
• For error-free transfers, MLED reached a maximum goodput of about 810 Mbps with no appreciable increase in delay.

Performance Results

The following results summarize the behavior of MLED and the traditional two-layer design for a 20,480 MB file under different PDU error rates and an adversarial error model. In this setup, errors are introduced in protocol data units (PDUs) in a way that allows them to evade the CRC and Internet checksum checks used by the traditional approach. As a result, the baseline design can deliver a corrupted final file, which then requires retransmission of the entire file after file-level integrity verification. At 0.000% error rate, both approaches perform similarly, reaching roughly 810 Mbps goodput and completing the transfer in about 200 seconds. As the error rate increases, however, the difference becomes pronounced.

MLED consistently achieves error-free transfers under this adversarial model, sustaining goodput close to 790–810 Mbps and file delivery time around 200–210 seconds. In contrast, the traditional approach drops to roughly 400 Mbps goodput and requires about 400–415 seconds to complete the transfer because corruption propagates to the final file and forces a full retransmission. These results show that MLED both preserves throughput and cuts recovery time nearly in half by detecting and correcting errors inside the network rather than relying on end-to-end detection after the full file has already been transferred.

Goodput comparison for MLED and the traditional approach across three PDU error rates. At non-zero error rates, MLED sustains close to 800 Mbps while the traditional design drops to about 400 Mbps.

File delivery time comparison for MLED and the traditional approach. Under non-zero error rates, MLED completes the transfer in about 200–210 seconds, while the traditional approach takes roughly 400–415 seconds.

Undergraduate Student Researchers

Ethan Frink

Arizona State University

Noah Barnes

Boston University

MLED Publications

Resources

This work is supported in part by NSF grants CNS-2215671 and CNS-2215672.

If you have any questions or would like to discuss the MLED framework further, please feel free to contact me.