SmartNIC-Enabled Live Migration for Storage-Optimized VMs with PYROCUMULUS
summary
Pyrocumulus offloads live migration of storage-optimized VMs into an FPGA SmartNIC running ~16K lines of SystemVerilog on an Intel Arria 10, attached to Intel 900P SSDs over a 100 G network, and implements a strong-consistency double-write protocol in hardware. The prototype reduces user latency variance by up to 12.4×, cuts migration time up to 19.6×, and lowers migration cost up to 3× while consuming only ~0.9% compute and 3.8% memory on the SmartNIC, eliminating large host CPU reservations.
AI-written summary · about
chapters (13)
- 0:00SmartNIC live migration intro
- 1:44Why live migration matters
- 3:10Storage VM limits
- 4:37Slow costly migration
- 7:24Parity Cumulus overview
- 8:46Data path with SmartNIC
- 10:11Data consistency in SmartNIC
- 11:36Prototype setup details
- 12:22Performance results
- 13:31Cost overhead analysis
- 14:33User impact and interference
- 15:40Future considerations
- 16:24Summary
related talks
18mIn-Memory Key-Value Store Live Migration with NetMigrate
01h 00mCommSec D1 - An Isolated Data Center Security Policy Model Using SmartNICs
18mNICA: An Infrastructure for Inline Acceleration of Network Applications
19mFpgaNIC: An FPGA-based Versatile 100Gb SmartNIC for GPUs
18mSTYX: Exploiting SmartNIC Capability to Reduce Datacenter Memory Tax
17mSRNIC: A Scalable Architecture for RDMA NICs
15mThe nanoPU: A Nanosecond Network Stack for Datacenters
17mOSMOSIS: Enabling Multi-Tenancy in Datacenter SmartNICs