Tracking Millions of Flows In High Speed Networks for Application Identification
-
Upload
amethyst-day -
Category
Documents
-
view
34 -
download
0
description
Transcript of Tracking Millions of Flows In High Speed Networks for Application Identification
![Page 1: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/1.jpg)
1
Tracking Millions of Flows In High Speed Networks for Application Identification
Tian Pan, Xiaoyu Guo, Chenhui Zhang, Junchen Jiang, Hao Wu and Bin Liut Tsinghua National Laboratory for Information Science and Technology Department of Computer Science and Technology, Tsinghua University, Beijing 10084, China
2012 Proceedings IEEE INFOCOM
通訊所 周啓松
![Page 2: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/2.jpg)
2
Outline
Framework of systemExternal flow table managementALFE replacement policyTheoretical analysisEvaluationConclusion
![Page 3: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/3.jpg)
3
FRAMEWORK OF SYSTEM
![Page 4: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/4.jpg)
4
Emerge issues
On-chip SRAM is fast but insufficient to accommodate millions of concurrent flows which has to be stored in DDR or RLDRAM.
High hash collision rate and high cache miss rate.
It is hard to attach parallel off-chip DRAMs to gain performance improvement.
![Page 5: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/5.jpg)
5
Purpose
An on-chip/off-chip hierarchical flow table to achieve high speed packet lookup and maintain tens of millions of flows.
Adaptive Least Frequently Evicted(ALFE), which keeps the elephant flows longer in the cache thus increasing the cache hit rate.
An efficient management scheme is proposed to exploit DRAM's burst feature.
• Real trace evaluation on Altera FPGA platform indicates, with 200MHz internal clock, small sized cache with 16K entries can achieve up to 80% hit rate, enabling more than 70Mpps line rate flow tracking even at the 40-byte packets.
![Page 6: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/6.jpg)
6
Framework of the Proposed Application Identification
![Page 7: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/7.jpg)
7
Design Trade-offs and System-level Optimization
Stateless TCP Handling without Flow Reconstruction◦ 3-Way shaking
◦ FIN/RST
Fixed-allocated Hash Buckets to Exploit DRAM Bursts◦ Lower utilization of memory space
Cache Replacement Policy to Track Elephant Flows◦ Heavy-tailed distribution
lmpact of Misidentification
![Page 8: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/8.jpg)
8
EXTERNAL FLOW TABLE MANAGEMENT
![Page 9: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/9.jpg)
9
Flow Record Data Structure
![Page 10: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/10.jpg)
10
Fixed-allocated Hash Bucket How to store the flow records How to read/write the memory efficiently
![Page 11: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/11.jpg)
11
Device Independent Layer
![Page 12: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/12.jpg)
12
Memory Swapping Logic
![Page 13: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/13.jpg)
13
ALFE REPLACEMENT POLICY
![Page 14: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/14.jpg)
14
Heavy-tailedness
![Page 15: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/15.jpg)
15
Replacement Policy
![Page 16: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/16.jpg)
16
THEORETICAL ANALYSIS
![Page 17: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/17.jpg)
17
Poisson Queuing Network System
![Page 18: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/18.jpg)
18
Throughput and Queue Length
The arrival rate of packets is less than the service rate:
The overall maximum throughput is determined by the maximum arrival rate which could be afforded by each part.
According to queuing theory, the system queue length is:
𝑄=(1+𝛼+𝛼𝛽)𝜆
𝜇𝑐−(1+𝛼+𝛼𝛽)𝜆+
(𝛼+𝛼𝛽)𝜆𝜇 𝑓 −(𝛼+𝛼𝛽)𝜆
![Page 19: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/19.jpg)
19
EVALUATION
![Page 20: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/20.jpg)
20
Experiment Setup
![Page 21: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/21.jpg)
21
On-chip Flow Cache
![Page 22: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/22.jpg)
22
Off-chip Flow Table
The collision rate will be bounded less than when active flow number (AFN) is 4M.
![Page 23: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/23.jpg)
23
Bucket OverflowNE
BAW
![Page 24: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/24.jpg)
24
Memory Utilization
![Page 25: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/25.jpg)
25
Delay and Throughput
The 400MHz DRAM I/O interface transfers two data words per DRAM clock cycle.
The read latency is 6 DCs. The write latency is 7 DCs. The maximum throughput of DRAM is 44.44Mpps.
𝜇𝐷𝑅𝐴𝑀≤400𝑀𝐻𝑧
7+⌈𝑁𝐸
2⌉
(𝑝𝑝𝑠 ) ,𝑁𝐸≥ 4
![Page 26: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/26.jpg)
26
Trade-off between Utilization and Speed
When BAW is set to be 15 bits, the system reach optimal balance. In this paper, the BAW is expanded to 21 bits to provide higher
speed.
◦ NE = 8
◦ Total entry number = 16.8M
![Page 27: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/27.jpg)
27
Power
![Page 28: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/28.jpg)
28
System Overall Performance System Throughput
◦ Overall throughput under different cache entry number
![Page 29: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/29.jpg)
29
avalanche effect
![Page 30: Tracking Millions of Flows In High Speed Networks for Application Identification](https://reader036.fdocuments.net/reader036/viewer/2022062720/568134dc550346895d9c0d98/html5/thumbnails/30.jpg)
30
Conclusion
Detail of Matching Engine should be listed.
Setup time of flow table should be taken into account System Overall Performance.