At the start of his PhD, Ahmed led an in-depth study of 136 network-partitioning failures across 25 widely used distributed systems, including file systems, messaging systems, and in-memory data structures. Notably, this work challenged key assumptions about network failures. For example, he found that 88% of failures can occur when a single node is isolated, urging network practitioners to prioritize top-of-rack switch failures, which were overlooked. Moreover, many practitioners assumed that eliminating client access would lower failures. However, his team found that most failures involved minimal to no client access. The team concluded their paper by introducing the Network Partitioning Testing (NEAT) framework, an open-source testing system.

This research, published at the USENIX Symposium on Operating Systems Design and Implementation, was widely praised for its industrial impact. For example, his team found flaws in the design of five common techniques, including scheduling, discovery service, and replication. They also identified a vulnerability to catastrophic failure in a common Zookeeper deployment approach and malpractice in Microsoft’s network maintenance procedures that aggravates the impact of network faults. Beyond the industry, this research was taught at several top-tier universities, among them Harvard University, the University of Michigan, and Johns Hopkins University. 

Ahmed also identified major bottlenecks in state-of-the-art replicated storage systems, which store copies of data across multiple machines to keep services available even when some machines fail. He and his co-researchers found that state-of-the-art designs struggle to fully leverage multicore machines. A key limitation is their reliance on a log of operations to coordinate replicas and track and order every operation in the system. This log becomes a major scalability bottleneck, limiting throughput and increasing latency.

This research culminated in LoLKV, the first logless and linearizable storage system. Instead of logs, LoLKV relies on per-key operation tracing to achieve linearizability. Ahmed and his collaborators have found that LoLKV has almost ten times higher throughput and up to 92% lower tail latency than other state-of-the-art storage systems. These findings were published at the 2024 USENIX Symposium on Networked Systems Design and Implementation.

Ahmed capped off his doctoral research by building efficient and reliable cloud computing systems. For example, he developed DROPS, a resource management system for serverless functions. DROPS uses a novel statistical analysis of historical execution traces to allocate resources efficiently while meeting performance requirements. Tech giant Microsoft Azure has adopted and deployed DROPS in their production, serving thousands of clients across the globe and handling millions of requests per day. It has even reduced Azure’s infrastructure costs by tens of millions of dollars. This research was published at the 21st European Conference on Computer Systems.

Next is Slogger, a new disaster recovery system that leverages modern data centre time synchronization hardware and protocols. It can reduce the data loss window by 50% compared with the traditional approach, incremental snapshotting, which only captures and stores data changes since the last backup. Ahmed created Slogger in collaboration with Oracle, who patented it.

Overall, Ahmed’s research helped advance and transform cloud computing systems, one of the major cornerstones of modern-day technology. From preventing network failures, including system crashes, to enhancing data recovery systems like phone back-ups, Ahmed’s lifelong work can help millions of users, right at their fingertips.