Towards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
summary
Pallas partitions rack traffic at the ToR so each shard has homogeneous CPU demand, uses a P4-programmed ToR switch for classification and per-group JSQ-style balancing, and assigns simple intra-server schedulers per shard. Implementation on real hardware shows in-network workload shaping cuts tail latency by 8.5× at medium load and up to two orders of magnitude at high load, raises sustainable throughput (up to 2.1× in some tests), and handles bursts via selective cloning.
AI-written summary · about
chapters (12)
- 0:00PARS introduction and goals
- 1:23Rack-scale μs scheduling challenges
- 3:53Proactive workload homogeneity
- 5:19PARS design: workload shaping
- 6:41Balancing homogeneity and efficiency
- 7:25Adapting to long-term workload changes
- 9:55Mitigating short-term bursts
- 11:19System overview and hardware
- 12:01Experimental evaluation
- 13:44Adapting to dynamic shifts
- 14:29Comparison with other methods
- 15:12Summary and open-source
related talks
25mVirtualizing Traffic Shapers for Practical Resource Allocation
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs
25mAvoiding Overload in Distributed Systems
18mHorus: Granular In-Network Task Scheduler for Cloud Datacenters
21mHandling Traffic Spikes Through Automated Prioritized Load Shedding
25mShenango: Achieving High CPU Efficiency for Latency-sensitive Datacenter Workloads
19mNet-P4ct: Enhanced WAN Bandwidth Fair Sharing Using P4 Programmable Switches
21mR2P2: Making RPCs first-class datacenter citizens