FreeBSD Network Stack Performance

Srinivas KrishnanUniversity of North Carolina at

Chapel Hill

Outline

Introduction Unix network stack improvements Bottlenecks

Memory Copies Interrupt Processing

Zero Copy Implementation Receive Live Lock Solution

Introduction

NIC

IP Queue

Packet

Transport + Network

Socket Queue

SoftInterrupt

Memory Copy

Memory Copy

Kernel

Processing

User

Processing

Network Stack Reinvented

Van Jacobson Net Channels Create a High Speed Channel from

NIC to User space Push all processing to the user space Applying E2E “truly” Preserve cache coherency for multi-

processor systems

BETTER INTERRUPT PROCESSING

Network Stack Reinvented

Ulrich Drepper’s Asynchronous Network I/O Asynchronous sockets True Zero Copy No Locking Event Channels

BETTER MEMORY PROCESSING

Reduce Memory Copies

Sending Side Copy from User Buffer to Kernel Buffer Copy from Kernel Buffer to Device Buffer

Receive Side Copy from Device Buffer to Kernel Buffer Copy from Kernel Buffer to Socket Buffer

Zero Copy Send

write

Userspace Pages

RAM

Page Sized chunks

External

mbuf

DMA into Driver Buffer

NIC

Zero Copy Read

NIC

Packet

Kernel BufferDMA

Kernel Space

User Space User Buffer

read(fd, buf, s)

Zero Copy

Allocate an External Mbuf Pool NIC MTU has to be >= 4K Intel Pro1000 NIC with Jumbo

Frames 3Com NIC turn on DMA

Buffer and stitch the data together Added Overhead

Page Flipping

Check Mbuf len

Page Size

! 1 PageUse

copyout

Usevm_pgmoveco (……)

Kernel Page <->User Pageread(….)

Atleast 1 Page

Preliminary Results

CPU Utilizatio (Kernel)

0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160 180 200

Data Rate (Mbps)

CP

U U

tiliza

tion (%

)

No Zero Copy

w Zero Copy

• 1500 bytes MTU (Iostat trace) for 10 mins

Processing Interrupts

Main Processing Hard Interrupt from NIC to driver Soft Interrupt from IP Queue to

processing Reduce user level and interrupt

thread processing Problem: Receive Live Locks

Receive Live Lock

Send large stream of UDP packets > receiver buffer capacity

CPU spent processing network packets

Goodput = 0

Implementation Design

NIC

IP Queue

Packet

Transport + Network

Socket Queue

Driver Queue

Scheduler

Components All UDP packets are queued in

driver queue Scheduler is triggered with the

arrival of first UDP packet Checks the queue every n ms

(currently 1-2ms) Schedules packet departure rate

based on timestamps

Driver Queue Algorithm

Set maximum rate and average rates

Driver Queue maintains Average Queue Length (Weighted

over time) Current Rate of transfer Time stamp of packets

Algorithm (cont) If current_rate > average rate

Drop N packets such that current_rate == average_rate

If current_rate > max rate (Spike) Drop all packets

Reduce Time Wait in Queue If Current Queue Size < threshold

Schedule packet exit such that rate == average_rate

Appends an exit time to each packet

Pros and Cons

Easy implementation requires no scheduling changes

Reduces CPU utilization in worst case by ~25%

Low Overhead Introduces added jitter

Experimental Setup

Receive UDP Data

Intel Pro1000 Nics

Send UDP Data

Intel Pro1000 Nics

•Iostat Trace

•Netstat trace

•Custom queue stats

Queue Stats

At the Receiver Collect Average Queue Size CPU Utilization Packet Drops Total Number of packets processed

Receive Live Lock

Receive Live Lock (soln)

Future Work

Feedback from Socket Queue and IP queue such that Weighted Average computed over all 3 queues

Drop at driver before DMA Driver buffer not large enough to keep

weighted queue size Feedback from Driver Queue Scheduler

to driver to drop

Questions ?