On the Performance Characteristics of High-Speed Enterprise Networks

(Performance Analysis of Underutilized Enterprise Network)

Goal: Identify the performance characteristics of today's high-speed enterprise-scale networks

Our primary analysis tool is to use measured TCP connection behavior at key observation points to draw scientifically grounded inferences on the factors leading to degraded network performance

1. We need to learn a bit more about POSTECH’s network
    1.1 Find and represent POSTECH network topology
      - POSTECH 캠퍼스 LAN 구성도, 스위치 설치 현황, 캠퍼스 코어 백본 스위치 port 현황에 대하여 전산소로부터 받은 자료를 바탕으로 현황 파악 (login required)
      - POSTECH weather map 구현을 통하여 Topology 조사 결과 반영 (LINK) (login required)
      - POSTECH Pajek: image (LINK), .net file (LINK)

   1.2 Perform SNMP measurement (coarse time scale)
      - POSTECH 캠퍼스 네트워크 자료를 기반으로 장비들 간의 연결정보 파악, SNMP 폴링을 통해 weather map에서 트래픽의 흐름 및 각 링크별 MRTG 그래프 확인 가능

   1.3 Asymmetric Link Analysis
      - Data caputure and analysis summary: (LINK)
           - Core 라우터의 하루 동안의 in/out 양: (LINK)
           - 인터넷 라우터 구간의 교내 호스트들의 트래픽 통계: (LINK)
           - 특정 서브넷별 layer 4 포트 번호별 트래픽 양: (LINK)

2. We need to check the soundness of the SNMP and DAG measurement system
   2.1 For SNMP, the main performance problem will be when fine granular (5-10 second) probing is carried out at a specific target switch at which a measurement PC is connected
      - SNMP Test Result Summary: (LINK)
      - 장비에 따라 MIB 값의 update 주기가 다름 (기숙사 메인 스위치가 가장 성능이 안 좋음, 8~9초)

   2.2 For DAG measurement, IP and TCP header and some X bytes of payload must be saved
    + Most important is to objectively test the performance range of the existing DAG measurement software system
      - DAG Test Result Summary: (LINK)
      - 약 300,000 pps 정도를 limit로 보고 있음

   2.3 More required test
      - Capture traffic using tcpdump for 1 hour under artificial test traffic
      - Examine that the inter-packet time stamps concur with the physical rate
      - Check how much space a 1 hour trace consumes

3. We select target switches to measure and perform both DAG measurements on specific links and SNMP measurements on all ports
   3.1 The DAG already in place at the Internet gateway will still be useful as is, with another placed at a primary/secondary switch determined after step (1)
      - POSTECH Monitoring Points 선정(LINK) (login required)

4. We should have clean performance data, especially TCP connection behavior, along with port-based SNMP data

Conference Meeting
  • 2007-03-09 (ppt)(pdf)
  • 2007-03-23 (pdf)
    Related Papers
  • Jitendra Padhye, Sally Floyd, "On Inferring TCP Behavior," ACM SIGCOMM 2001, San Deigo, California, USA, August 2001.

     Comments to pluto80[at]postech.ac.kr   (last updated March 08, 2007)