Research paper Measured Capacity of an Ethernet: Myths and Reality

Size: px

Start display at page:

Download "Research paper Measured Capacity of an Ethernet: Myths and Reality"

Jeffry Barrett
5 years ago
Views:

1 Research paper Measured apacity of an Ethernet: Myths and Reality Theoretical work seems to suggest that Ethernet works saturate at 7%. Realistic networks can offer higher throughputs Lessons learnt Don t install long cables: to cover a large area, break up the cable with bridges or gateways (routers), not repeaters. Don t put too many hosts on one cable: use gateways to break the network into communities of interest, trading higher delay for inter-community traffic for better intra-community response time and throughput urrent ethernets are aset or aset and use switches Feb-4-4/598N: omputer Networks Feb-4-4/598N: omputer Networks Lessons learnt Implement the protocol correctly: proper collision detection and binary exponential backoff in interface or host software is essential to good performance Use the largest possible packet size: this keeps the packet count down, reducing the likelihood of collision and not incidentally reducing overheads internal to hosts Especially important for Gigabit Ethernets. They define Jumbo frames (9K packets). More on this for HWP Outline Switching and Forwarding Store-and-Forward Switches ridges and Extended LNs ell Switching Segmentation and Reassembly Don t mix serious real-time and serious bulk-data applications: it is not possible to simultaneously guarantee the lowest delay and the highest throughput (although for moderate requirements both kinds of applications coexist well) Feb-4-4/598N: omputer Networks Feb-4-4/598N: omputer Networks 4 Switch Scalable Networks forwards packets from input port to output port port selected based on address in packet header T T STS- Input ports Switch T T STS- Output ports dvantages cover large geographic area (tolerate latency) support large numbers of hosts (scalable bandwidth) Switch Host Source Routing Switch Switch Host Feb-4-4/598N: omputer Networks 5 Feb-4-4/598N: omputer Networks 6

2 Virtual ircuit Switching Explicit connection setup (and tear-down) phase Subsequence packets follow same circuit Sometimes called connection-oriented model nalogy: phone call 5 Host Switch Switch Virtual ircuit Model Typically wait full RTT for connection setup before sending first data packet. While the connection request contains the full address for destination, each data packet contains only a small identifier, making the per-packet header overhead small. If a switch or a link in a connection fails, the connection is broken and a new one needs to be established. Each switch maintains a V table 7 Switch 4 Host onnection setup provides an opportunity to reserve resources. Feb-4-4/598N: omputer Networks 7 Feb-4-4/598N: omputer Networks 8 Datagram Switching No connection setup phase Each packet forwarded independently Sometimes called connectionless model nalogy: postal system Each switch maintains a forwarding (routing) table Host Host D Host Switch Host E Switch Host G Switch Host Host F Datagram Model There is no round trip time delay waiting for connection setup; a host can send data as soon as it is ready. Source host has no way of knowing if the network is capable of delivering a packet or if the destination host is even up. Since packets are treated independently, it is possible to route around link and node failures. Since every packet must carry the full address of the destination, the overhead per packet is higher than for the connection-oriented model. Host H Feb-4-4/598N: omputer Networks 9 Feb-4-4/598N: omputer Networks ridges and Extended LNs LNs have physical limitations (e.g., 5m) onnect two or more LNs with a bridge accept and forward strategy level connection (does not add packet header) Ethernet Switch = ridge on Steroids ridge Port Port X Y Z Learning ridges Do not forward when unnecessary Maintain forwarding table ridge Port Port X Y Z Host Learn table entries based on source address Table is an optimization; need not be complete lways forward broadcast frames X Y Z Port Feb-4-4/598N: omputer Networks Feb-4-4/598N: omputer Networks

3 Spanning Tree lgorithm Problem: loops - no mechanism to remove looping frames E G I 6 D F H J K ridges run a distributed spanning tree algorithm select which bridges actively forward developed by Radia Perlman now IEEE 8. specification lgorithm Overview Each bridge has unique id (e.g.,,, ) Select bridge with smallest id as root Select bridge on each LN closest to root as designated bridge (use id to break ties) Each bridge forwards frames over each LN for which it is the designated bridge E G I 6 D F H J K Feb-4-4/598N: omputer Networks Feb-4-4/598N: omputer Networks 4 lgorithm Details ridges exchange configuration id for bridge sending the message id for what the sending bridge believes to be root bridge distance (hops) from sending bridge to root bridge Each bridge records current best configuration message for each port Initially, each bridge believes it is the root Feb-4-4/598N: omputer Networks 5 lgorithm Detail (cont) When learn not root, stop generating config in steady state, only root generates configuration When learn not designated bridge, stop forwarding config in steady state, only designated bridges forward config Root continues to periodically send config If any bridge does not receive config message after a period of time, it starts generating config claiming to be the root Feb-4-4/598N: omputer Networks 6 roadcast and Multicast Forward all broadcast/multicast frames current practice Learn when no group members downstream ccomplished by having each member of group G send a frame to bridge multicast address with G in source field Do not scale Limitations of ridges spanning tree algorithm does not scale - traffic gets bridged through the root bridge Spanning tree is designed to avoid loops, not traffic balancing: redundant routes are ignored broadcast does not scale Do not accommodate heterogeneity aution: beware of transparency Feb-4-4/598N: omputer Networks 7 Feb-4-4/598N: omputer Networks 8

4 Smartridges se_smbr_k.asx 78 Hybrid between IP routing and bridging ell Switching (TM) onnection-oriented packet-switched network Used in both WN and LN settings Signaling (connection setup) Protocol: Q.9 Specified by TM forum Packets are called cells 5-byte header + 48-byte payload ommonly transmitted over SONET other physical layers possible Feb-4-4/598N: omputer Networks 9 Feb-4-4/598N: omputer Networks Variable vs Fixed-Length Packets No Optimal Length if small: high header-to-data overhead if large: low utilization for small Fixed-Length Easier to Switch in Hardware simpler enables parallelism ig vs Small Packets Small Improves Queue behavior finer-grained pre-emption point for scheduling link maximum packet = 4K link speed = Mbps transmission time = 496 x 8/ = 7.68us high priority packet may sit in the queue 7.68us in contrast, 5 x 8/ = 4.4us for TM near cut-through behavior two 4K packets arrive at same time link idle for 7.68us while both arrive at end of 7.68us, still have 8K to transmit in contrast, can transmit first cell after 4.4us at end of 7.68us, just over 4K left in queue Feb-4-4/598N: omputer Networks Feb-4-4/598N: omputer Networks ig vs Small (cont) Small Improves Latency (for voice) voice digitally encoded at 64Kps (8-bit samples at 8KHz) need full cell s worth of samples before sending cell example: -byte cells implies 5ms per cell (too long) smaller latency implies no need for echo cancellors TM ompromise: 48 bytes = (+64)/ User-Network Interface (UNI) GF ell Format VPI VI Type LP HE (R-8) host-to-switch format GF: Generic Flow ontrol (still being defined) VI: Virtual ircuit Identifier VPI: Virtual Path Identifier Type: management, congestion control, L5 (later) LPL ell Loss Priority HE: Header Error heck (R-8) 8 84 (48 bytes) Payload Network-Network Interface (NNI) switch-to-switch format GF becomes part of VPI field Feb-4-4/598N: omputer Networks Feb-4-4/598N: omputer Networks 4 4

5 Segmentation and Reassembly TM daptation Layer (L) L and designed for applications that need guaranteed rate (e.g., voice, video) L /4 designed for packet data L 5 is an alternative standard for packet data L L L /4 onvergence Sublayer Protocol Data Unit (S- PDU) < 64 K PI tag Size User data Pad Etag Len PI: commerce part indicator (version field) tag/etag:beginning and ending tag size: hint on amount of buffer space to allocate Length: size of whole PDU TM TM Feb-4-4/598N: omputer Networks 5 Feb-4-4/598N: omputer Networks 6 ell Format Type OM: beginning of message OM: continuation of message EOM end of message SEQ: sequence of number MID: message id Length: number of bytes of PDU in this cell (44 bytes) 6 TM header Type SEQ MID Payload Length R- S-PDU Format L5 < 64 K 47 bytes 6 6 Data Pad Reserved R- pad so trailer always falls at end of TM cell Length: size of PDU (data only) R- (detects missing or misordered cells) ell Format end-of-pdu bit in Type field of TM header Len Feb-4-4/598N: omputer Networks 7 Feb-4-4/598N: omputer Networks 8 5

Switching and Forwarding

Switching and Forwarding Outline Store-and-Forward Switches ridges and Extended LNs ell Switching Segmentation and Reassembly Scalable Networks Switch forwards packets from input port to output port port