Questions That Led Somewhere Interesting

“Questions that started with a performance anomaly, an unexpected result, or a systems problem and ended up teaching me something useful.”

E-WarPDRAMReal-Time SystemsBenchmarking

How do you generate predictable DRAM contention for an experiment?

Stress-testing memory bandwidth is easy. Stress-testing it the same, specific, worst-case way on every single run is a different problem entirely — and the one I actually had to solve to validate E-WarP.

7 min readRead the investigation →

E-WarPMemGuardReal-Time SystemsResearch Process

Why we abandoned phase detection for cumulative memory envelopes?

A signal-processing approach for breaking an application into memory-behavior stages turned out to be the wrong abstraction for predicting MemGuard stalls.

9 min readRead the investigation →

Cloud SystemsResource ManagementProPerE-WarP

Why do we keep relearning the same application every time it runs?

Most resource managers watch an application, react to what it's doing, and forget everything the moment it exits. For software that runs the same code paths a million times, that isn't caution. It's waste.

8 min readRead the investigation →

Intel RDTCache PartitioningMemory BandwidthReal-Time Systems

Why I had to reverse engineer Intel instead of reading the manual?

Embedded platforms come with an explicit hardware contract. Intel's resource-management interfaces looked similar on paper — until experiments across two processor generations showed that the manuals described an interface, not necessarily the behavior that mattered for predictable systems.

9 min readRead the investigation →