A Practical Packet Reordering Mechanism

1
A Practical Packet Reordering Mechanism with
Flow Granularity for Parallel Exploiting in
Network Processors
13th WPDRTS April 4, 2005
Beibei Wu, Yang Xu, Bin Liu, Hongbin
Lu Department of Computer Science, Tsinghua
University, Beijing, P.R.China
2
Background Problem

• Network Processor (NP)
• A special purpose, programmable hardware device
  that combines the flexibility of a RISC processor
  with the speed of ASIC. They are building blocks
  used to construct network systems
• Data path Processing Engine (PE)
• Two Design Goals
• High Speed Multiple PEs
• packet level
  parallelism
• High Flexibility Versatile processing
  requirements
• unpredictable
  processing time for each packet
• The Packet Disordering (PD) Problem
• Packets depart in a different order from their
  arrival
• Network performance may be deteriorated greatly

3
Objective
NP Model
PE0
Dispatcher
Aggregator
PEn

• To design a practical mechanism which can
  preserve packet order in NP, at the same time to
  ensure the utilization of PE

4
Contents

• Design Issues
• Global-scope Vs. within-flow-scope order
  preserving
• Pre-processing Vs. post-processing order
  scheduling
• The Proposed Solution
• Packet chains for all the flow sequence
  information
• The working process
• Simulation

5
Design Issues(1)the scope of packets to preserve
order

• Global-scope
• All packets leave strictly in order
• Within-flow-scope
• Only packets of the same flow leave in order
• Processing delay of different flows is different
  in NP
• within-flow-order-preserving is quite necessary

6
Design Issues(2)Where order scheduling is taking
place?

• Order scheduling location
• Pre-processing scheduling
• SPSL--Sequential Processing Sequential
  Leaving
• Post-processing scheduling
• UPSL-- Un-sequential Processing Sequential
  Leaving

7
Design Issues(2)the shortcoming of
pre-processing scheduling?
PE0
Dispatcher
Aggregator
PEn
Packet Buffer
8
Contents

• Design Issues
• Global-scope Vs. within-flow-scope order
  preserving
• Pre-processing Vs. post-processing order
  scheduling
• The Proposed Solution
• Packet chains for sequence information
• The working process
• Simulation

9
The Proposed Solution (1)The methods of
reordering

• in traditional network devices
• Sequence number or timestamp
• global order preserving
• in our NP
• Packet chains
• Providing ability to discriminate among flows

10
The Proposed Solution (2)NP System with Packet
Data Buffer
PE Complex
t1
p2
pn
p1
t2
tm
b1
b2
bk
Dispatcher
Packet Data Buffer
Aggregator

• packet lt-gt thread lt-gt block
• blocks lt-gt disordered packets buffering

11
The Proposed Solution (3)

• Packet in the NP lt-gt block in the Packet Data
  Memory
• When to transmit packet in which block?
• How to discriminate among flows?

12
The Proposed Solution (4) using packet chains to
record sequence information
Flow Table
Block Table
f(1)
p(a)
p(b)
f(j)
p(e)
FlowID
End Packet
Head Packet
p(d)
p(x)
p(y)
p(z)
FlowID
packet
Dispatcher
13
The Proposed Solution (5)The Working Process
PE 1
Packet Data Buffer
f
4
b1
PE 2
r
1
f
b2
f
5
b3
f
b4
1
2
3
4
5
r
2
b5
f
f
b6
f
3
b7
f
b8
f0
5
7
1
5
3
7
2
flowID
head
end
14
Contents

• Design Issues
• Global-scope Vs. within-flow-scope order
  preserving
• Pre-processing Vs. post-processing order
  scheduling
• The Proposed Solution
• Packet chains for sequence information
• The working process
• Simulation

15
Simulationthroughput vs flow number, for three
traces
A NP system with 4 PEs, each 4 threads and
totally 8432 memory blocks

• Trace1 f1 100 constant 40bytes f2 0
• Trace2 f1 95 constant 40bytes f2 5
  random from 40 to 60bytes
• Trace3 f1 90 constant 40bytes f2 10
  random from 80 to 120bytes

f1 is length unrelated app. Packets,
f2 is length related app. packets
Where
16
Simulation Utilization of 4 PEs under traces
with the fewest flows

• Time fraction of active thread number for each PE

trace1
trace2
trace3
All traces have the fewest flows
17
Simulation Buffer Occupation with the fewest
flows

• Time fraction of free block number for each PE

trace1
trace2
trace3
All traces have the fewest flows
18
A Summary

• A solution to preserve packet order in NP with
  multiple PE for data plane processing
• Packet chains to record sequence information of
  different flows to preserve packet order
• Reordering within-flow-scope is quite necessary
  in NP
• Future work optimize Memory block and PE
  resources

19
Thank you for your attention!