CSE221 - lec17: Networking: IX & Snap

Datacenters

• applications must be scalable to run on multiple machines

• high network bandwidth: 100 - 400 Gbits/s (~120 ns per packet)

• storage: high speed flash, NVM

• must minimize overhead: kernel crossing is too expensive. (since many operations are so fast and may be frequent)

Kernel Bypassing

• traditional network stack: packet processing is in the kernel, the applications need to do sys call to get or send the packets.

• kernel bypassing: have a library that could access the NIC directly and application could access packets via RAM. (eliminate kernel crossings)

Data path v.s. Control path

• data path: sending / receiving packets

• control path: management tasks

Isolation in modern NICs

• separate virtual queues

• IOMMU: allows NIC to translate virtual address

IX: A Protected Dataplane Operating System for High Throughput and Low Latency

Challenges

• High packet rates

• resource efficiency

• protection: between applications, network stack is trusted

• microsecond tail latency: strict latency requirement, usually we are talking about things like p99 latency

IX Overview

Run to completion & Adaptive batching

• run to completion: process a packet from start to finish

• adaptive batching: dynamically vary the batch size based on load

Zero copy

• POSIX API: copy packets (from kernel to user / user to kernel)

• zero copy: design api to avoid copy
• drawback: hard programming model; may be harmed by uncooperative program

Synchronization free processing

• partition work across cores

• each core use dedicated NIC queues
• Receive Side Scaling
• hash packets to different queues

Summary

• protected data plane using hardware virtualization

• run to completion, adaptive batching, RSS, zero copy

Snap: a Microkernel Approach to Host Networking

Goals (besides those of IX)

• developing velocity

• maintainability, flexibility

• easier deployment

• easier optimization

Structure of Snap

LibOS Approach v.s. Microkernel Approach

Threading

• use kernel threads

IPC

• lower overhead due to multi-cores, IPC can be done in different cores

Snap Microkernel

• control v.s. data plane

• engines: basic execution units of network stack processing

• cpu scheduling:
• dedicated mode: dedicating several cores to snap exclusively
• spreading engines: spread snap threads to all cores
• compact: only let snap compete for a small number of cores

MicroQuanta Kernel Scheduling Class

• guarantee snap could run for certain portion of time in a period

Transparent Upgrade

• the blackout period (when service is down) is about 250 ms, and snap allows weekly upgrade.

Summary

• microkernel based approach for easier updates

• different modes for engines

• widely used in google