tcpwirespeed_an_tfs_tm_ae.pdf

7/29/2019 tcpwirespeed_an_tfs_tm_ae.pdf

1/8

Application Note

WEBSITE:www.jdsu.com/test

TCP Wirespeed: Testing TCP Throughput to 10 G

Managing TCP/IP-based Applications in the Ethernet World

As Ethernet speeds rise rom 1Gb/s to 10 Gigabit Ethernet (GigE), networks continue to grow increasinglycomplex. raditionally, the network provider has been responsible or Layers 1-3 o the network OpenNetwork Initiative (OSI) stack while the end-users I department manages has been responsible or theapplication layers (5-7). Reer to Figure 1 which illustrates the various layers o the OSI Model.

Te gray area suraces in the management o Layer 4 or ransmission Control Protocol/Internet Protocol(CP/IP)-based applications such as HP, FP, and e-mail. Network providers must prove that more thanjust the Layer 3, IP network is unctioning properly because IP perormance can be excellent, while CPperormance can be poor. Improperly tuned CP causes degradation in application perormance and end-

customers oen blame the network. Te problem may be in the network providers domain (lack randomearly discard queuing, improper prioritization, etc.) or within the end customers domain (i.e., poorly tunedCP window, lack o CPE trac shaping, etc).

Figure 1: OSI Model

Figure 2: Workow or Application-Centric Turn-Up

o remain competitive, network providers must continue to evolve their test capabilities as new tech-nologies emerge. Tose who can test Carrier Ethernet and IP eld networks using application-level testsor Layer 4 (L4) throughput will dramatically reduce nger-pointing between the end-customer andthe network provider. Figure 2 highlights an enhanced workow or turning up L2/L3-based IP services,emphasizing application/service-level verication.

Physical1

Datalink2

Network3

Transport

4

Session5

Presentation6

Responsibility

Customers IT

Department

Shared

Network Providers

HTTP, FTP, E-mail, etc.

TCP

IP

Ethernet

Application7

Traditional L2/L3 Turn-Up

Connectivity

Throughput (RFC-2544)

Class of Service Verification

Voice, Video, and Data CoS

Application Layer Turn-Up

TCP, UDP

FTP

HTTP


2/8

Application Note: TCP Wirespeed: Testing TCP Throughput to 10 G 2

Providers must be equipped with a eld communication test tool that can test various CP parametersand optimize CP perormance against network conditions such as latency. Improper CP congurationcoupled with network latency is one o the biggest culprits resulting in poor application perormance. Mostapplications ride on top o CP, such as FP and HP, whose perormance is directly related to properCP perormance. Providers must test CP rst, tune the CP conguration, and then test FP and HP.

The Fundamentals of TCP Window Sizing

Te CP window size reers to the amount o data bytes that a sender will place onto the network beore anacknowledgment is required rom the receiver. In the example in Figure 3, the CP window is ull aersending 17 kB which occurs in about 3 ms; no more data can be sent till the ACK is received.

Figure 3. Improper TCP Settings Reduce Application Throughput

o urther illustrate, improper CP settings on a 3 link with 25 ms round-trip delay (RD) can easilyreduce application throughput by 80 percent (6 vs 45 Mb/s). Conducting tests with multiple windowsizes, to nd one that is ideal or maximum throughput, simplies tuning and testing eforts and lets usersquickly determine i low CP throughput was the result o improper window conguration.

Te primary concept or CP throughput testing involves the bandwidth delay product (BDP). With Layer3 tests (below the CP layer), the latency o the network does not afect the achievable Layer 3 throughput.In other words, a 45 Mb/s link with 25 ms delay could achieve nearly 45 Mb/s at Layer 3.

But or CP (Layer 4), the latency is key along with the CP window size setting. Again, the CP windowsize is the amount o data bytes that a sender will place onto the network beore an acknowledgment is

required rom the receiver.

Te BDP is the product o a data links capacity (in bits per second) and its end-to-end delay(in seconds).Te equation or the BDP is as ollows:

BDP = RTD * Link Speed

Te BDP calculation provides the CP window as the ideal number o bits, dividing by 8 will calculate theideal CP window. In the example o 45 Mb/s link and 25 ms round-trip delay, a window size o ~140 kBis optimal. For rame sizes o 1514 bytes, this equates to about 95 rames sent to the receiver beore a CPacknowledgment is required.

Sender with

Window=17K

Receiver

ACKACK takes 12.5 ms to reach Sender

T3 Link with 25 ms RTD, (latency=12.5 ms)

Internet

#1 #1 #1 #12 *sending stops
http://en.wikipedia.org/wiki/Multiplicationhttp://en.wikipedia.org/wiki/Bit_ratehttp://en.wikipedia.org/wiki/End-to-end_delayhttp://en.wikipedia.org/wiki/End-to-end_delayhttp://en.wikipedia.org/wiki/Bit_ratehttp://en.wikipedia.org/wiki/Multiplication


3/8


The JDSU TCP Wirespeed Option: Simplifying the Complexity of TCP TestingTe JDSU -BERD/MS-6000A supports stateul CP sessions at a line rate o up to 10G and cansupport up to 64 CP sessions. Another benet o the -BERD/MS-6000A is that it is compatible withthe commonly used iperfsoware tool, but eliminates the limitations and inaccuracies o PC-based iperperormance.

As mentioned in the previous section, the CP window must be properly tuned with the latency o thenetwork in order to achieve ull throughput. Te establishment o the proper window size can be a com-plicated and tedious task, but the JDSU -BERD/MS-6000A Walk the Window script automates theprocess to allow less experienced technicians to conduct the tests.

Aer activating the Walk the Window script, the user simply congures the starting and ending windowsizes, the number o windows to test (between the starting and ending window size), and the duration oeach trial test, as Figure 4 shows.

Figure 4. TCP Walk the Window Script Conguration

Te test is then run between the two end-end JDSU -BERD/MS-6000 or a 6000A and an iper server.Aer completion, the user is presented with an easy-to-understand test report that clearly highlights thewindow size versus throughput perormance, as Figure 5 shows. For this test, the link speed was 1 GigEand the RD was 1 ms. Tis example shows that the optimum CP window size is 128 kB which achievesmaximum Layer 4 throughput o ~950 Mb/s.


4/8


Figure 5. TCP Walk the Window Script Results

Aer establishing the optimum window size, CP throughput testing should be conducted with a mixtureo background trac such as UDP to veriy customer throughput under realistic network conditions.Router queue settings commonly cause customer throughput issues and cannot be diagnosed unless anetwork is under load. CP ail Drop issues are surprisingly widespread in networks and many times goundetected until a customer exceeds a particular trac threshold.

TCP Tail Drop Explained

In properly congured routers, the bufer mechanism should be set to periodically drop packets when the

router becomes congested. Routers use various Early Discard algorithms to alleviate congestive situations.

I a router is congured improperly, the bufer will ll and cause the router to drop the tail o the bufer.

This drop causes a drop o a sequential block o packets that can adversely efect the built-in retransmission

mechanisms o TCP. The net result is that the multitude o TCP hosts (such as servers and PCs, or example)

begin to retransmit in synchronicity, which can cause an endless loop o router congestive overload and

discards. The user then experiences an incredible slowdown in TCP sessions such as HTTP and, in many

cases, is the cause o the long spinning hourglass on Windows workstations.


5/8


Automating the TCP Test ProcessEven though RFC2544 was meant to benchmark network equipment and to be used by networkequipment manuacturers (NEMs), network providers also use it to benchmark operational networksin order to provide Service Level Agreements (SLAs) to their business customers. Ultimately, networkproviders realize that a successul RFC2544 test result does not guarantee end-user satisaction, thusrequiring, an automated means or testing CP perormance.

Network providers (and NEMs) are wrestling with end-to-end complexities and must standardize themethodology to validate end-to-end CP perormance, as this is the precursor to acceptable end-userapplication perormance. JDSU is developing a standard methodology or testing CP layer perormanceand guidelines or expected CP throughput results or observation in the network under test.

Te next subsections highlight the important elements o the JDSU CP testing methodology.

End-to-End TCP-Only Installation Testing

Te -BERD/MS-6000A CP Wirespeed option, when used in the conguration shown in Figure 6,provides results that characterize the achieved CP throughput at various CP window size settings. Inaddition, this tester provides resultant retransmission counts (due to packet loss) and ofers much greatercondence or providers that customer applications will achieve expected throughput.

Figure 6. T-BERD/MTS-6000A acting as both the TCP Client and TCP Server

End-to-End TCP and Streams Installation Testing

Figure 7 shows how operators can veriy proper CoS/QoS prioritization in the network by conducting anend-to-end test with concurrent CP and background streams to prove that customer application sessions(CP) will receive expected priority versus background trac types such as UDP streams.

P

T-BERD/MTS-6000A

TCP Client

T-BERD/MTS-6000A

TCP Client

P

T-BERD/MTS-6000A

TCP Client

T-BERD/MTS-6000A

TCP Client

64 TCP Sessions up to 10G

4 Background Streams up to 10G

Figure 7. Using the T-BERD/MTS-6000A to Veriy Proper CoS/QoS Prioritization using Background Streams


6/8


TCP Throughput Testing with an iperf HostFigure 8 shows the -BERD/MS-6000A MSAM acting as a Client communicating with an iper Serverrunning industry-standard, open source CP test soware. Tis test lets network operators go one stepurther in terms o throughput segmentation by either using a test server or the actual customer serverto conduct CP throughput tests. Te -BERD/MS-6000A can be considered a golden data sourcebecause its hardware CP implementation transmits CP data at the ull line rate.

Figure 8. T-BERD/MTS-6000A MSAM as the TCP Client Communicating with an iper Server

Ciscos ttcp mode

Most Cisco devices support the ttcp TCP throughput utility which is also compatible with the T-BERD/

MTS-6000A. The 6000A test unit can be used to sectionalize TCP perormance across Cisco devices as it is alsocompatible with ttcp.

P

T-BERD/MTS-6000ATCP Client

Iperf Server(running on PC)

Up to 64 TCP Sessions


7/8


ConclusionAs wire speeds continue to rise and new applications emerge, networks increase in complexity. Fornetwork owners to remain competitive, they must evolve the turn-up approach as new technologiesemerge. Network operators must conduct turn-up tests that emulate actual customer trac in orderto prove that their networks deliver SLAs at the CP layer, which will eliminate nger-pointing whenproblems arise.

Applications such as le downloads, e-mail, and Internet access can perorm poorly when CP is runningpoorly. L4 tests are required to validate and ensure network SLA and oen require the expertise o anexpert network engineer. However, employing the use o the -BERD/MS-6000A can equip a ier 1technician to conduct CP throughput tests at line rates up to 10G that can support as many as 64 CPsessions. It is compatible with the industry-standard iper soware tool which allows the -BERD/MS-6000A to be used interoperably with iper-based client and servers.

Te -BERD/MS-6000A can be used to veriy proper CoS/QoS prioritization, conduct CP throughputtests, or to automatically characterize CP perormance at various CP window sizes. Network operatorsand service providers who can veriy perormance accurately and ensure SLAs can reduce customerchurn and eliminate nger-pointing when issues arise.


8/8


Productspecicationsand descriptionsin this documentsubject to changewithoutnotice. 2010 JDS UniphaseCorporation 30162985 000 0210 CPWIRESPEE D.AN.FS.M.AE February 2010

Test & Measurement Regional Sales

NORTH AMERICATEL: 1 866 228 3762

FAX: +1 301 353 9216

LATIN AMERICATEL: +1 954 688 5660

FAX: +1 954 345 4668

ASIA PACIFICTEL: +852 2892 0990

FAX: +852 2892 0770

EMEATEL: +49 7121 86 2222

FAX: +49 7121 86 1222

WEBSITE: www.jdsu.com/test

tcpwirespeed_an_tfs_tm_ae.pdf

Documents

Transcript of tcpwirespeed_an_tfs_tm_ae.pdf