1. Low remote block I/O throughput

What is the current status?

• Traditional iSCSI protocol requires:
  - 14 CPU cores to saturate a 1M IOPS SSD.
  - 56 CPU cores to saturate a 100Gbps link.
• User-level approaches require:
  - Changes in applications and/or networks.
• NVMe-over-Fabrics:
  - (RDMA) Changes network infrastructure.
  - (TCP) Suffers low performance.

Technical challenge: “Can we push back the performance bottlenecks to the hardware, requiring no modifications in applications/networks?”

2. Multi-tenancy support

How to support various tenants’ requirements?

Each host/target server can have multiple tenants with different target devices, requirements, resources, and so on.

• (As-Is) The current storage I/O stack creates per-core queues, and each I/O request takes a static data path to the target device on each core.
• (To-Be) The request should be able to change its data path dynamically at low cost, in order to satisfy all tenants’ requirements (e.g., throughput, latency, etc.).

Technical challenge: “What would be the best abstraction to support multi-tenancy at block device layer?”