arXiv subjects

Georgia Fragkouli

Publications and source records attributed to Georgia Fragkouli.

3 recordsLinked to original sources

arXiv2025

Faster Offloads by Unloading them -- The RDMA Case

From hardware offloads like RDMA to software ones like eBPF, offloads are everywhere and their value is in performance. However, there is evidence that fully offloading -- even when feasible -- does not always give the expected speedups. Starting from the observation that this is due to changes the offloads make -- by moving tasks from the application/CPU closer to the network/link layer -- we argue that to further accelerate offloads, we need to make offloads reversible by unloading them -- moving back part of the offloaded tasks. Unloading comes with a set of challenges that we start answering in this paper by focusing on (offloaded) RDMA writes: which part of the write operation does it make sense to unload? how do we dynamically decide which writes to execute on the unload or offload path to improve performance? how do we maintain compatibility between the two paths? Our current prototype shows the potential of unloading by accelerating RDMA writes by up to 31%.

cs.NI↗

arXiv2025

The Effects of iBGP Convergence

Analyzing violations of forwarding properties is a classic networking problem. However, existing work is either tailored to the steady state -- and not to transient states during iBGP convergence -- or does analyze transient violations but with inaccurate proxies, like control-plane convergence, or without precise control over the different impact factors. We address this gap with a measurement framework that controllably and accurately measures transient violation times in realistic network deployments. The framework relies on a programmable switch to flexibly emulate diverse topologies and gain traffic visibility at all links -- enabling accurately inferring violation times of any forwarding property. Using the framework, we analyze 50 network scenarios on a topology with 12 real routers, and show how factors like the network configuration and BGP event affect transient violation times. Further, we shed light on less-known aspects of BGP convergence, including that transient violations can start before the trigger event, or that keeping a backup route advertised at all times can increase violation times.

cs.NI↗

arXiv2022

Limiting Lamport Exposure to Distant Failures in Globally-Managed Distributed Systems

Globalized computing infrastructures offer the convenience and elasticity of globally managed objects and services, but lack the resilience to distant failures that localized infrastructures such as private clouds provide. Providing both global management and resilience to distant failures, however, poses a fundamental problem for configuration services: How to discover a possibly migratory, strongly-consistent service/object in a globalized infrastructure without dependencies on globalized state? Limix is the first metadata configuration service that addresses this problem. With Limix, global strongly-consistent data-plane services and objects are insulated from remote gray failures by ensuring that the definitive, strongly-consistent metadata for any object is always confined to the same region as the object itself. Limix guarantees availability bounds: any user can continue accessing any strongly consistent object that matters to the user located at distance $Δ$ away, insulated from failures outside a small multiple of $Δ$. We built a Limix metadata service based on CockroachDB. Our experiments on Internet-like networks and on AWS, using realistic trace-driven workloads, show that Limix enables global management and significantly improves availability over the state-of-the-art.

cs.DC↗