NetPerL Seminar

1
NetPerL Seminar

• An Analysis of TCP Processing Overhead
• David D. Clark, Van Jacobson, John Romkey, Howard
  Salwen
• IEEE Communications Magazine, June 1989

Presented By Jim Harris July 23, 2002
2
Background

• Classic paper notable authors, citations
• Issue performance of TCP protocol
• Scientific approach to address issue
• Seminal conclusions on TCP/IP performance that
  have withstood the passage of time

3
The Analysis

• Started with TCP in BSD UNIX, and modified to
  remove dependencies on UNIX
• Determined common path for TCP, and implemented a
  fast path
• See figure

4
Analysis Terminology
5
A First Case Study Input Processing

• Common belief that receiving is more complex
  paper refuted belief
• Three stages of receive processing
• Search for TCB
• TCP checksum
• Packet header processed
• First, study packet header processing
• Counted instructions

6
TCP Output Processing / IP

• TCP output processing
• Sending (output) side less detailed analysis
• Used instruction count again
• IP processing
• Instruction estimates for both receiving and
  sending common path
• Header prediction example
• IP block copy of header template

7
Support Functions / Checksums and TCBs

• The buffer layer
• Buffer layers for protocols have same goal
• Specifications of protocol suite intentionally
  broad in this area
• Timers and Schedulers
• Timer code can be cost demanding
• Checksums and TCBs
• Issue of CPU or outboard chip for checksum
• Temporal coherency for TCB blocks observed

8
Some Speed Predictions

• Receiving a packet estimated at 300 instructions
  for TCP/IP
• Used factor of 4/3 for RISC to get 400
• Used 100Mbp/s FDDI and 10 MIPS processor with one
  ack for every 2 packets yields a packet sized of
  500 bytes before protocol is the limiting factor,
  i.e., bottleneck

9
Why Are Protocols Slow?

• Real sources of overhead
• Operating system interrupt processing, buffer
  management, context switches (scheduling), timer
  operation
• Byte touching operations moving data in memory,
  checksumming
• Data copy biggest factor CPU and DMA

10
A Direct Measure of Protocol Overhead

• Used Sun-3/60 to measure time with logic analyzer
  our HP16700 has this capabilty

11
A Direct Measure of Protocol Overhead (cont)

• Times yield estimated instruction counts that are
  consistent with above analysis
• 70 of time in TCP, and 80 of that in output
  (send) side - counter to common belief
• Times support OS impact
• Per byte operations swamp protocol processing
  costs 771 us versus 100 us
• Protocol processing small compared to OS
  processing 240 us versus 100 us

12
Some Dangerous Speculations

• If OS and memory overhead the real limits, then
  move the processing to a special outboard
  processor
• Problems identified with buffer layer between
  host and outboard processor
• Alternative design of simple network interface
  hardware minimize interrupts, minimize data
  copies, compute checksum with data copy

13
Protocols in Silicon

• Not necessary to abandon TCP/IP protocol to
  improve performance
• Recommend moving to
• Efficient network interface card, or
• Simple network controller on cpu bus
• Desirable to have a programmable element using
  general protocols
• TCP has 15 years of improvements which will
  continue as of 1989, now TCP/IP are at least 28
  years old
• Putting protocols in hardware is not required