Name
Abstract | ||
|---|---|---|
Elastic computing such as disaggregated storage systems increasingly adopts RDMA to accelerate data plane operations (e.g., get and put). However, such computing systems usually require ephemeral operations to scale out computing resources, which demands fast control plane operations such as starting containers and establishing network connections. Unfortunately, building a connection for RDMA usually requires several or tens of milliseconds, which is 180-1800X higher than the execution time of typical RDMA-enabled systems (10-100 {\mu}s) requiring elasticity. Such a high control plane cost may significantly reduce the efficiency of such systems when scaling out the computing resources. This paper presents KRCORE, an RDMA library with a microsecond-scale control plane on commodity RDMA-capable network interface cards. KRCORE can establish a network connection to any node in the cluster within 10{\mu}s, while only maintaining a (small) fixed size of connection metadata at each node, regardless of the cluster scale. The key ideas include virtualizing a pre-initialized kernel-space RDMA connection instead of creating one from scratch, and retrofiting advanced RDMA dynamic connected transport with static transport for low connection overhead and high resource utilization. KRCORE shortens the worker bootstrap time of existing disaggregated key-value store by 98% under load spikes. In serverless computing, another popular elastic and ephemeral computing system, KRCORE can reduce the latency for transferring data through RDMA by 99%. |
Year | Venue | DocType |
|---|---|---|
2022 | USENIX Annual Technical Conference (USENIX ATC) | Conference |
Citations | PageRank | References |
0 | 0.34 | 0 |
Authors | ||
4 | ||
Name | Order | Citations | PageRank |
|---|---|---|---|
| Xingda Wei | 1 | 42 | 4.20 |
| Fangming Lu | 2 | 0 | 0.34 |
| Rong Chen | 3 | 586 | 30.22 |
| Haibo Chen | 4 | 1749 | 123.40 |