Analysis of Network Processing Workloads

1
Analysis of Network Processing Workloads
Authors Ramaswamy Ramaswamy, Ning Weng, and
Tilman Wolf Publisher IEEE 2005 Present
Shih-Chin Chang (???) Date Thursday, December
31, 2009
Department of Computer Science and Information
Engineering National Cheng Kung University,
Taiwan
2
Outline

• Introduction
• PacketBench
• PacketBench System
• PacketBench API
• PacketBench Prototype
• Application Workload
• Network Processing Applications
• Network Traces and Routing Tables
• Results
• Average Statistics
• Variation across Packets
• Individual Packet Analysis

3
Introduction
4
Introduction (cont.)

• Compared to other network processing simulators
  and benchmarks, PacketBench
• First, be programmed easily
• Second, allow applications to operate on actual
  packets in the same fashion as it is done inside
  the NP.
• Third, be able to hide the overhead for packet
  preprocessing in the PacketBench framework.
• PacketBench can be used in
• Application Optimization
• Allocation of Processing Tasks
• Developing Novel NP Architectures
• Understanding the Dynamics of Network Processing

5
Outline

• Introduction
• PacketBench
• PacketBench System
• PacketBench API
• PacketBench Prototype
• Application Workload
• Network Processing Applications
• Network Traces and Routing Tables
• Results
• Average Statistics
• Variation across Packets
• Individual Packet Analysis

6
PacketBench
PacketBench Framework
7
PacketBench System

• The key point about this system design
• The application and the framework can be clearly
  distinguished even though both components need
  to be compiled into a single executable in order
  to be simulated.
• This is done by analyzing the instruction
  addresses and sequence of API calls.

8
PacketBench System (cont.)

• PacketBench Framework
• To provide functions that are necessary to read
  and write packets, and manage memory.
• PacketBench API
• To distinguish between PacketBench and
  application operations during simulation.
• Network Processing Application
• To implement the actual processing of the
  packets.
• Processor Simulator
• To get instruction-level workload statistics.

9
PacketBench API

• To define how applications can receive, send, or
  drop packets.
• void init()
• This function is implemented by the application
  and called by the framework before any packets
  are processed.
• void (process_packet_function)(packet )
• This function is the packet handler that is
  implemented by the application.
• void write_packet_to_file(packet , int)
• This function is implemented by the framework and
  called by the application when processing is
  complete.

10
PacketBench Prototype

• Use the ARM target of the SimpleScalar simulator
  to analyze the applications.
• The ARM architecture is very similar to the
  architecture of the core processor and the
  microengines found in the Intel IXP 2400 network
  processor.
• The tools were setup to work on an Intel x86
  workstation running RedHat Linux 9.0.
• PacketBench supports packet traces in the tcpdump
  format and the Time Sequenced Header (TSH) format
  from NLANR.

11
Outline

• Introduction
• PacketBench
• PacketBench System
• PacketBench API
• PacketBench Prototype
• Application Workload
• Network Processing Applications
• Network Traces and Routing Tables
• Results
• Average Statistics
• Variation across Packets
• Individual Packet Analysis

12
Application Workload

• Network Processing Applications
• IPv4-radix
• IPv4-trie
• Flow Classification
• TSA (Top-hashed Subtree-replicated Anonymization)
• Network Traces and Routing Tables

13
Outline

• Introduction
• PacketBench
• PacketBench System
• PacketBench API
• PacketBench Prototype
• Application Workload
• Network Processing Applications
• Network Traces and Routing Tables
• Results
• Average Statistics
• Variation across Packets
• Individual Packet Analysis

14
Results

• Three classes of results that can be derived
• Microarchitectural Results
• Network Processing Results
• Per-Packet Analysis
• These statistics are divided into three
  categories
• Averaged Statistics
• Variation across Packets
• Individual Packet Analysis
• Processing complexity the average number of
  instructions executed per packet.
• The memory accesses and coverage statistics can
• Distinguish between instruction and data memory.
• Separate data memory into packet data and program
  data.

15
Average Statistics

• Processing Complexity

The overhead of maintaining and traversing the
radix tree.
Take the same amount of instructions.
16
Average Statistics (cont.)

• Memory Accesses
• Accesses to packet memory (e.g. the packet
  header, the payload)
• Accesses to non-packet memory (e.g. routing
  tables, flow tables)
• This analysis was performed for the first 10,000
  packets of each trace.

17
Average Statistics (cont.)

• Memory Coverage
• By SimpleScalar
• This analysis was performed on the first 1,000
  packets of the MRA trace.

Due to large data structures.
18
Variation across Packets

• Processing Complexity

The first 500 packets of the MRA trace.
The first 100,000 packets of the COS trace.
19
Variation across Packets (cont.)

• Memory Accesses

20
Individual Packet Analysis

• Instruction Pattern

21
Individual Packet Analysis (cont.)

• Unique Instructions

The first 100,000 packets of the COS trace.
22
Individual Packet Analysis (cont.)

• Basic Block Access Frequency

23
Individual Packet Analysis (cont.)

• Basic Block Coverage

24
Individual Packet Analysis (cont.)

• Memory Access Sequence