Presentation is loading. Please wait.

Presentation is loading. Please wait.

 SNU INC Lab 2015-10-05 고속 packet processing 기술 김 종 권김 종 권.

Similar presentations


Presentation on theme: " SNU INC Lab 2015-10-05 고속 packet processing 기술 김 종 권김 종 권."— Presentation transcript:

1  SNU INC Lab 2015-10-05 고속 packet processing 기술 김 종 권김 종 권

2  SNU INC Lab 2015-10-05 Contents  Introduction Forwarding engine 현재의 IP 패킷 처리 방법  Router 에의 QoS 기능 추가 Packet filtering(classification) Scheduling algorithms  Open problems  Summary

3  SNU INC Lab 2015-10-05 Common Architecture

4  SNU INC Lab 2015-10-05 FE 의 기능  Input IP header Inbound MAC header 의 일부  Output MAC header info Outbound interface Outbound next hop  Performance FE module 의 성능 = 전체 system 의 성능 Pipelining  한 packet 의 header 를 처리하는 동안 다음 packet 의 header load

5  SNU INC Lab 2015-10-05 FE 알고리즘  병렬화  문제점 Route lookup 후에만 TTL = 1/0 인 packet 을 폐기한다. Route lookup 후에만 header 를 갱신하다. 1. Error Check Lengths Version Martians Checksum TTL 2. Route Lookup Cache Full 3. Header Update TTL Checksum Link Layer header Fragmentation IP 전송의 세 부분

6  SNU INC Lab 2015-10-05 Router 의 IP Packet 처리 방법  Packet 도착 Link level processing (CRC check 등 )  Link 정보를 IP 에 제공  Header 확인 IP header checksum, version, header length 등 확인  주소 확인 Link-layer 주소와 의미가 일치하는 지 확인  Router 를 위한 packet 결정 Router 를 위한 packet 은 다르게 처리된다.  Router 를 위한 packet 의 목적지 주소

7  SNU INC Lab 2015-10-05 Router 의 구조 Fig. Integrated services architecture implemented in router

8  SNU INC Lab 2015-10-05 Queueing Disciplines  Packet filtering(Classification) Map each arriving packet into one of classes Proposed algorithm (SIGCOMM ‘98)  “Fast and Scalable Layer Four Switching”  Washington University  V. Srinivasan, G. Varghese, S. Suri, M. Waldvogel  “High Speed Policy-based Packet Forwarding Using Efficient Multi-dimensional Range Matching”  Bell Labs  T. V. Lakshman, D. Stiliadis  Scheduling 어떤 순서로 packet 들을 서비스 할 것인가 ? Desired link sharing 고려

9  SNU INC Lab 2015-10-05 Wash Univ Algo (1/ 3)  Grid-of-tries Two-dimensional filters

10  SNU INC Lab 2015-10-05 Wash Univ Algo (2/ 3)  Cross-Producting 각 Field 별로 따로 검색하여, cross-product table 의 hash table 을 lookup Memory explosion problem

11  SNU INC Lab 2015-10-05 Wash Univ Algo (3/ 3)  Comparison Grid-of-tries  Destination-source prefix filter 에만 적용 가능 On-demand cross-producting  임의의 filter 에 적용 가능  Cache miss 때문에 speed 를 예상하기 어렵다.  Grid-of-tries 와 Cross-producting 의 혼합 Grid-of-tries  Destination-source filters On-demand cross-producting  나머지 filters

12  SNU INC Lab 2015-10-05 Bell-Lab Algo (1/ 3)  Point location problem in multidimensional space Given a point in a d-dimensional space, and a set of n d- dimensional objects,  Preprocessing 각 dimension 마다 생기는 interval 을 계산한다. 각 interval 을 포함하는 filter 들을 계산한다.  Packet classification Packet 의 각 차원 별로 어느 interval 에 속하는지 계산한다. 각 interval 을 포함하는 filter 들의 교집합을 구한다. 가장 우선 순위가 높은 filter 를 사용한다.

13  SNU INC Lab 2015-10-05 Bell-Lab Algo (2/ 3)  2-dimension 인 경우의 예

14  SNU INC Lab 2015-10-05 Bell-Lab Algo (3/ 3)  Hardware Implementation Integer comparator, counter AND operation n 번의 binary search n 개의 값에 대한 AND 연산

15  SNU INC Lab 2015-10-05 Scheduling  필요성

16  SNU INC Lab 2015-10-05 Weighted Fair Queueing  특징 GPS simulation(Packet GPS) Variable-size packets  Algorithm GPS 를 simulation 하여 그 결과에 따라 service 순서 결정 Finish number  GPS 에서 packet 의 finishing time

17  SNU INC Lab 2015-10-05 WFQ - Evaluation  Evaluation GPS scheduling 에 비해 maximum packet size 보다는 늦게 전송되지 않는다는 것을 보장 Isolation 보장  최대 지연은 다른 connection 의 behavior 에 영향 받지 않는다.  Problem Expensive and complex to implement O(N)

18  SNU INC Lab 2015-10-05 Virtual Clock  Real time 사용  Time complexity Virtual time 계산 : O(1) Sorting: O(log N)  단점 Unbounded unfairness

19  SNU INC Lab 2015-10-05 Self-Clocked Fair Queueing  현재 서비스 중인 packet 의 finish number 를 system 의 virtual time 으로 사용.  Time complexity Virtual time 계산 : O(1) Sorting: O(log N)  단점 End-to-end delay bound  N Unfair over short time scales

20  SNU INC Lab 2015-10-05 Weighted Round-Robin  Weighted round-robin Different weight Different size  Divide each connection’s weight by its mean packet size to obtain a normalized set of weights.  단점 Source’s packet size may be unpredictable.  Source cannot predict its mean packet size. Fair only over time scales longer than a round time.  If a connection has a small weight, or the number of connections is large, this may lead to long periods of unfairness.  Good in high speed network like ATM

21  SNU INC Lab 2015-10-05 Deficit Round-Robin  DRR can handle variable packet size without knowing the mean packet size. 1500 800 1200 20001000 Second Round First Round Head of Queue A B C 0 Quantum size : 1000 byte  1st Round A’s count : 1000 B’s count : 200 (served) C’s count : 1000  2nd Round A’s count : 500 (served) B’s count : 0 C’s count : 800 (served)

22  SNU INC Lab 2015-10-05 Frame-based Fair Queueing  Potential function 사용  Re-calibration periodically  Fairness 보장 Bounded period system potential update Bounded discrepancy

23  SNU INC Lab 2015-10-05 Leap Forward Virtual Clock  두개의 priority queue 사용 H: Priority Queue of Active Well-behaved Flows L: Holding Area for Active Oversubscribed Flows Transfer before a packet can miss its deadline. Transfer when a flow becomes oversubscribed. Service Packet with Lowest tag in H

24  SNU INC Lab 2015-10-05 LFVC Implementation & Evaluation  Implementation van Emde Boas Data Structure 사용 O(loglog N)  Evaluation Data Structure 가 H/W 로 구현하기에는 너무 복잡하다. Worst case 에는 L 에서 H 로 옮겨야 하는 packet 이 N 개일 수 있다.

25  SNU INC Lab 2015-10-05 Discrete-rate Approach  Per-rate queueing for ATM 같은 rate 를 갖는 session 들을 하나로 묶어 macro-session 형성  Two-level hierarchical scheduler First level: GPS-related scheduler  각 macro-session 의 output-link access 관리 Second level: Round-robin fashion  각 session 에 rate 분배

26  SNU INC Lab 2015-10-05 H/W Implementation  Calendar queue 사용

27  SNU INC Lab 2015-10-05 Core-Stateless Fair Queueing  Core-Stateless architecture Edge nodes: do perform per flow management Core nodes: do not perform per flow management

28  SNU INC Lab 2015-10-05 CSFQ Algorithm  Edge router Maintain per flow state Estimate the incoming rate Label the packet  Core router Maintain no per flow state FIFO packet scheduling with probabilistic dropping algorithm Drop each packet with prob. P(= max{0, }) Update packet label to min{, }

29  SNU INC Lab 2015-10-05 Open Problems - by Keshav  Flow Identification Fast classification algorithm, efficient flow descriptors  Resource Reservation Efficient flow classifiers, good policing  Ease of Configuration Configuration tools  Stability of Large Systems Stability of router software  Accountability Pricing

30  SNU INC Lab 2015-10-05 Summary  Current research issues Table lookup QoS  Open problems

31  SNU INC Lab 2015-10-05 References  “Fast and Scalable Layer Four Switching”, V. Srinivasan, G. Varghese, S. Suri, M. Waldvogel, SIGCOMM’98.  “High Speed Policy-based Packet Forwarding Using Efficient Multi-dimensional Range Matching”, T. V. Lakshman, D. Stiliadis, SIGCOMM’98.  “Efficient Fair-Queueing Algorithms for Packet- Switched Networks”, D. Stiliadis, A. Varma, Trans. on Net., Apr. 1998.  “Hardware Implementation of Fair Queueing Algorithms for Asynchronous Transfer Mode Networks”, A. Varma, D. Stiliadis, IEEE Comm., Dec. 1997.

32  SNU INC Lab 2015-10-05 References  “Leap Forward Virtual Clock: A New Fair Queuing Scheme with Guaranteed Delays and Throughput Fairness ”, S. Suri, G. Varghese, G. Chandranmenon, INFOCOM’97.  “High Speed, Scalable, and Accurate Implementation of Packet Fair Queueing Algorithms in ATM Networks”, J. Bennett, D. Stephens, H. Zhang, ICNP’97.  “Implementing Fair Queueing in ATM Switches: The Discrete-Rate Approach”, F. Chiussi, Andrea Francini, INFOCOM’98.  “Core-Stateless Fair Queueing: Achieving Approximately Fair Bandwidth Allocations in High Speed Networks ”, I. Stoica, S. Shenker, H. Zhang, SIGCOMM’98.


Download ppt " SNU INC Lab 2015-10-05 고속 packet processing 기술 김 종 권김 종 권."

Similar presentations


Ads by Google