Hyrax: Fail-in-Place Server Operation in Cloud Platforms
summary
Hyrax lets servers continue hosting VMs with failed but degradable components by disabling faulty parts (using an uncommon rank-bitmask deactivation for DIMMs) and exposing per-page bandwidth to the hypervisor via page-coloring to equalize bandwidth. Evaluated on production Azure servers and large-scale simulations, Hyrax keeps peak VM performance within ~5% for 32-core VMs, cuts technician-required repairs by ~50–60% (55% median), lowers TCO ~4.5% for 6-year lifetimes, and slashes DIMM and SSD replacements.
AI-written summary · about
chapters (13)
- 0:00Hyrax project intro
- 1:56All-or-nothing limitations
- 3:00Goal: lower repair costs
- 4:44Hyrax workflow concept
- 6:09Hyrax implementation challenges
- 7:12Deactivate with firmware
- 8:40Memory deactivation impact
- 10:05QoS with page coloring
- 10:26Hyrax impact on VM scheduling
- 11:27Simulation results
- 12:29Repair frequency and TCO
- 13:11Hyrax additional benefits
- 13:51Future research
related talks
32mHVX: Virtualizing the Cloud
21mHybrid Cloud - nothing to worry about! 🌥️
48mAzure Stack HCI and Microsoft Hybrid Data Services
14mHyperscale Infrastructure Q&A
22mInfrastructure & the Cloud
01h 11mConnect your servers to the Cloud using Azure Hybrid Management
15mMake a Once-in-a-Generation Leap – Migrate to Azure Ampere-based VMs
29mMicrosoft Azure: Operating at Hyper-Scale