Bandwidth Measurements - University of California, Berkeley

Download Report

Transcript Bandwidth Measurements - University of California, Berkeley

Bandwidth Measurements

Jeng Lung WebTP Meeting 10/25/99

Bandwidth Definitions

 Bottleneck Bandwidth: The ideal bandwidth of the lowest bandwidth link (bottleneck link) on a route between two hosts. This quantity is independent of traffic.

 Available Bandwidth: Maximum bandwidth that a host can transmit at along a route.

 Alternative Definition:How fast a connection should transmit to preserve network stability.

Previous Work On Bottleneck Bandwidth Measurements

 Packet Pair  Probe Packets  Bprobe  PBM  Pathchar  ROPP & PBF Keshav 1991 Bolot 1993 Carter & Crovella 1996 Vern Paxson 1997 Van Jacobson 1997 Lai & Baker 1998

Available Bandwidth Measurements

 Available Bandwidth not well defined  Possible metrics include: – Throughput: Amount of data a transport protocol can transfer per unit time.

– TCP Congestion Avoidance: Send more and more packets until packet drop occurs.

Packet Pair

  If two packets are transmitted faster than the bottleneck link, then when they arrive at the bottleneck link they will be spread out in time by the transmission delay. If the spacing remains the same when the packet pair reaches the receiver, the bottleneck bandwidth can be inferred.

Assumptions: The second packet queues behind the first packet at the bottleneck link and that both packets travels on the same route. Also assume FIFO store and forward router queuing.

Packet Pair Bandwidth Estimation

 t transmit   t bottleneck   t arrive  Bandwidth = Packet Size / Separation

Difficulties with Packet Pair

  Competing traffic: – – Time compression: Other packet queue ahead of the first probe packet when it is downstream of the bottleneck link. This leads to high estimates.

Time extension: Other packet delay the second probe packet and extends the spacing between the two probe packets. This leads to low estimates.

Lack of queuing at bottleneck link: The probe packets were not sent fast enough to cause queuing at the bottleneck link. (Transmitting the packets slower than the bottleneck bandwidth would cause this)

More Difficulties with Packet Pair

     Packet drops Multiple routes: – Out of order packet delivery – Multi-channel bottleneck links Clock resolution: Can’t measure bandwidth higher than the one limited by the clock resolution.

Changing bottleneck bandwidth: Routing changes or ISDN channel activating a second channel.

Asymmetric Bandwidth: For methods that measures round trip time instead of one way transit time.

Probe Packets (Bolot)

 Use round trip delay of UDP echo packets to form estimates of bottleneck bandwidths.

 Stream of packets are sent at fixed intervals and the bottleneck bandwidth is calculated by comparing the round trip time of adjacent packets from a sequence.

Probe Packet Result

   The model assumes RTT(n+1) = RTT(n) + P/u – d X intercept at d – P/u (eg. d = transmission interval = 50 ms, P = packet size = 32 bytes. If x intercept is 48 ms, then u = bandwidth = 128 kb/s)

Bprobe (Carter and Crovella)

 Bottleneck bandwidth is estimated by transmitting different sets of 10 consecutive ICMP echo packets and recording the interarrival times of the consecutive replies. Each set has its own fixed packet size.

 The focus is on filtering the raw measurements in order to provide a solid estimate.

Bprobe

 Each run has a set probe packet size which increases from bottom to top in the graph. The estimates clusters more with increasing packet size which shows, in general, larger packet sizes yield more consistent results.

Bprobe Filtering

   Expand an error interval around each estimate and applying either intersection or union operation on the data set.

Intersection filtering: Find overlaps between estimate intervals and computes the intersection. (start with estimates from the largest packet set and iteratively intersect with estimates from successively smaller packets) Union filtering: Combines overlapping intervals using set union and pick the point at which most intervals overlap.

Union and Intersection Filtering

Bprobe Results

RBPP, SBPP (V. Paxson)

   Sender Based Packet Pair (SBPP) measures the roundtrip time at the sender. It has the problem of ACK compression.

Receiver Based Packet Pair (RBPP) measures the arrival time at the receiver and is more accurate than SBPP.

Time compression is detected when the difference between the transmission time of two packets is greater than the difference between their arrival times (RBPP) or their round trip times (SBPP).

PBM (V. Paxson)

 Form estimates for a range of packet bunch sizes and allow for multiple bottleneck values.

 Different bunch sizes can accommodate limited receiver clock resolutions and possibility of multiple channels.

 Bottleneck estimate is formed by clustering bandwidth estimates.

Pathchar (V. Jacobson)

 Able to measure the bandwidth of every link on a path.

 It is slow and can consume tremendous amounts of network bandwidth.

 Not scalable: sends 10 MB of data for a 10-hop measurement regardless of the bandwidth of the route.

Problems with current Packet Pair Algorithms

     Lack statistical robustness: Use of heuristics such as histogram or error intervals.

Do not use existing traffic: Send probe packets for measurements instead of using existing traffic.

Do not give a timely estimate: Estimate is calculated after series of probe packets or after a large sample.

Do not work on all traffic: Use of heuristics to handle small packets.

Lack flexibility to bandwidth changes.

Improvements to Packet Pair (Lai and Baker)

 Passive packet pair implementation which uses existing network traffic.

 Gradual algorithm which gives bandwidth estimate for every packet that arrives. A limited window of past packets is used to calculate bandwidth and to detect bandwidth changes.

 Filtering is done by using a kernel density instead of heuristics or histograms.

Packet Pair Filtering

      Filter out the noise caused by time compressed and time extended packets.

The main idea is that there is correlation among correct estimates (hence valid samples closely clustered around the correct value) and lack of correlation among incorrect estimates.

Find the point of greatest density in the distribution of bandwidth estimates.

The kernel function K(x) = {1+x (x  0), 1-x (x>0)} The density at any point x is 1

n i n

  1 1

h K

(

x

h xi

) (h is kernel width, n is the number of points within h of x, and xi is the ith such point.

Receiver Only Packet Pair (ROPP) (Lai and Baker)

 ROPP only take timing measurements from the receiver and hence it can’t filter out time compressed packets like SBPP or RBPP.

 It is much less likely than SBPP to have such samples because it is not relying on round trip time.

 Almost as accurate as RBPP but without having timing information from both sender and receiver.

Potential Bandwidth Filtering (Lai and Baker)

    Existing traffic maybe unsuitable for Packet Pair if the the packets are small or are sent slower than the bottleneck link bandwidth.

Potential bandwidth is the rate at which the sender sends the packets.

Filter out unsuitable samples by looking at the correlation between the potential bandwidth and the measured bandwidth.

Samples with high potential bandwidth and low measured bandwidth are the most informative.

Potential Bandwidth Filtering

x = y knee at x = b y = b Bandwidth Samples Potential Bandwidth

Simulation Results

 Use ns with 87 node network.

 ROPP achieves accuracy within 1% of RBPP.

 Potential Bandwidth Filtering is 37% to 435% more accurate on average than Measured Bandwidth Filtering in some cases.

Sources

      Srinivasan Keshav. A control-theoretic approach to flow control. In Proceedings of SIGCOMM, 1991.

Jean-Chrysostome Bolot. End-to-end packet delay and loss behavior in the internet. In Proceedings of SIGCOMM, 1993.

Robert L. Carter and Mark E. Crovella. Measuring bottleneck link speed in packet switched networks. Technical Report BU-CS-96 006, Boston University, 1996.

Vern Paxson. End-to-end internet packet dynamics. In Proceedings of SIGCOMM, 1997.

Van Jacobson. Pathchar. 1997.

Kevin Lai and Mary Baker. Measuring Bandwidth. 1998.