John S. Graham Global Research NOC Synopsis Network Performance Problems Definitions Network Testing Going Loopy Purpose Taxonomy Testing New Leased Lines Dedicated Test Sets Network Contribution to Poor Performance 1. Incorrect MTU Configuration 1. Standardized IP MTU of 9000 Bytes 2. Check end-to-end: ping -M do -c 1 -s 8972 <far end server> 2. Causes of Packet Loss in the Network 1. Due to a Congested Line 2. Due to a Line Introducing Transmission Errors 3. Process Switching by a Router in the Path iPerf Testing A UDP iPerf test will probe the network for packet loss Always employ a UDP test: Eliminate Layer-IV Flow Control from affecting results Simplex test selectively probes one direction (Source Sink or Sink Source) The bitrate can be operator controlled Use a modest packet rate for testing High-bitrate tests do not reveal additional information Excessive test traffic may cause congestion! Whos the Villain? MY Router Interface Or THEIR Circuit? 10G Circuit Over DWDM Ring Direct Router-to-Router Adjacency over 10GE Interfaces Kansas City Houston Miami Atlanta Representation of DWDM Circuit Tributary Side Line Side Cross-Connect Termination Points Tributary Side Line Side Soft Loop Classification Facility Loops Loop is applied to fiber side of CTP Equipment (aka Terminal) Loops Loop is applied to cross-connect side of CTP Supported Only on Tributary CTP Infinera DTN Four Possible Configurations Supported on Trib & Line CTPs Ciena CoreDirector Eight Possible Configurations A Bestiary of Loops Testing From a Router 1. Identify Far-End IP Address 2. Configure ARP Mapping : {IP} FE <-> {MAC} NE
3. Clear Egress Interface Counters 4. Send Measured Number of Pings to {IP} FE 5. EITHER Looped Network Clear of Errors 1. UCAST_IN = UCAST_OUT = Pings TTL 6. OR Looped Network Introduced Errors 1. UCAST_IN < UCAST_OUT Testing a New Leased Line 10G Leased Line iperf -c 192.168.8.1 -u -b10m Juniper MX960 Brocade MLX-16 A/32 B/32 192.168.8.1 A/32 192.168.8.1 B/32 4 2 1 3 Static Route Firewall Rule (Inbound) P2P VLAN A Typical Ethernet Tester 2 10G Ports (XFP Optics) Management Network Testers vs Software Probes Network Testers Do not process packet headers. Operator can design packets with any combination of protocols Header fields may contain any value Optimized for packet counting Will accept, identify and count every inbound frame; even those returned by a loop Software Probes Only accept packets with valid headers. Packet design severely restricted. Many header fields not accessible Accurate packet counting difficult with standard tools Unable to operate in conjunction with a looped network connection