VoIP & Enum VoIP & Enum PerformancePerformance

Charles Shen, Henning SchulzrinneDept. of Computer Science

Columbia UniversityNew York, NYMarch 30, 2006

Status UpdateStatus Update PDNS ENUM Server Performance

Instrumenting PDNS source code done Writing Performance Test Tool done Tests at Columbia in analysis phase Ready to port to Fibernet Phonomenun

infrastructure for tests VoIP/Enum call setup delay

Studied existing work on call setup delay in both PSTN and VoIP world

Ready for next step measurement

OutlineOutline ENUM query performance

Enumperf tool PDNS internal What’s being measured

PSTN/ISDN and VoIP call setup delay Targets values Expectations Case studies

Enumperf Enumperf Testing scripts we wrote, built on PDNS,

Nominum QueryPerf, and modeled from SIPstone.

Supply parameters through the configuration file, the scripts handle the tests and output result curves at the end.

Output parsing is currently specific to PDNS. Not difficult to adapt it to be used with other authoritative DNS servers.

PDNS Internal Thread ViewPDNS Internal Thread View

Main Thread

server control

listening

Distributor

N threads to serve queries

Query Listening

Pkt Cache

Query Cache

DB

Query input

Server control input

PDNS Query Processing PDNS Query Processing FlowFlow

Receive QueryCheck

Packet_ Cache

EnterQuestion Queue

Actual LookupLook up for

Canonic Record

Look up recursively for

subdomain SOA

Direct Match found

Direct Match found

Look up for NS records

SOA found

SendoutResponse

Y

N

Y

N Y

NServfail

NS foundANYfound

SOAfound

Look up for ANY records

Look up for SOA records

NN

YYY

Unauthoritative NS referralAuthoritative empty

NO ERROR

Authoritative NXDOMAIN

Authoritative empty NO ERROR

N

NO ERROR response

NO ERROR response

Things being investigatedThings being investigated Three types of queries

With matching records No matching records, but with subdomain SOA No matching records, no matching subdomain SOA

Bottom line performance (no caching) Caching effect

PDNS Packet Cache PDNS Query Cache MySQL Cache

Effect of number of threads Effect of database scaling Effect of Collocated vs. Separate MySQL/PDNS

Server Anything else?

Telephony Setup Delay Telephony Setup Delay DefinitionsDefinitions Post-dialing delay (aka post-selection delay)

interval between entering the last dialed digit and receiving ringback.

Dial-to-ring delay the interval between entering the last dialed

digit and when the callee's phone starts to ring. Answer-Signal Delay

interval between when callee picks-up the phone, and the time the caller receives indication of this.

Call Release Delay interval between when the releasing party

hangs-up and the time he can initiate/receive a new call.

Telephony Delay Target Telephony Delay Target E721E721

Telephony Delay Target Telephony Delay Target (Cont)(Cont)

Based on overall target values in ITU-T E721, E723 provides delay target values for SS7 Initial Addressing

Message (IAM) and Answer Message (ANM) partition for national and international portions

Total International

National

IAM 4.0 1.5 2.5

ANM 2.5 1.0 1.5

Intelligent Network Service Intelligent Network Service DelayDelay ITU-T E724 defines incremental post-

selection delay for processing IN services. Such as database query, call redirection,

multiple setup attempts, multiple processing locations, user interaction with the network, non-call related messaging.

12s 99th Percentile for processing all IN services in local connections.

4s mean delay for a single IN service for which the caller is unaware that IN processing is involve led (e.g. call forwarding)

Rest undefined.

Reference Connection ModelReference Connection Model Reference connection model correlates delays

incurred in each network component with E2E delay. E723 specifies basic reference connection for

NISDN. Q709 specifies a more detailed Hypothetical

Signaling Reference Connection model (HSRC) that: contains national components and international

components. international working consists of one international and two

national components. the size of the country roughly defined as large or average the respective number of nodes in each components.

E728 specifies the reference connection for BISDN signaling.

Delay Expectations from Delay Expectations from UsersUsers A no-longer-available ATT link said to have reported

that calls in US typically set up in 1-2 seconds. An expired Internet Draft estimates the international

call expectation to be roughly less than three times of the national call 2.5-5s. (based on maximum number of nodes in connection model)

A Nominum white paper says PSTN call establishment is expected to take less than 200 milliseconds.

ITU-T E724 describes mean user threshold for IN: processing all IN Services: 3.5s in local and 4.5s in toll

connections Single IN service 1.5s for caller unaware services and 1.0 s

for freephone-type services.

Computing Telephony Computing Telephony DelayDelay Cross Switch Delay

Q725 (processing intensive/simple messages); Telecordia spec (ISUP etc. Q931 message types); Q766 (IAM, ANM etc)

Cross STP Delay Processor Handling time - Q706 Outgoing link delay (queuing + emission) – Q706

Signaling link propagation delay – Q706 E2e - Reference connection Model, e.g. Q709 Call setup - Signaling exchange flow, SS7

Case Studies Case Studies Case I HSRC Q709

Used Q706 and Q725/Q766 maximum single message signaling delays

for simple/intensive message in each national component (0.3~0.4s), international component (0.3~0.5s) and e2e (range of 1s).

Adding Message specific delay values to get delays for specific message (rather than just processing simple/intensive). (ID ref)

Consideration of database query delay (Q709).

Case Studies (cont.)Case Studies (cont.) Case II BISDN Q728

Used Q706 and Q725. IAM (0.8/1.5/2.5) and ANM (0.5/1.0/1.8)

mean for local, toll, international. Case III PLMS Q771

The E721 delay targets were assumed for the fixed network

Added Authentication, paging/alerting, routing number transfer delays.

9-11-14 for F-M; 5.5-7.5-10.5 for M-F and 11.5-13.5-16.5 for M-M for local – toll – international connections.

Mapping Delay to VoIP Mapping Delay to VoIP worldworld

INVITE

100 Trying

ACK

BYE

200 OK

Media

180 Ringing

200 OK

Post Dialing Delay

Answer Signal Delay

Call Release Delay

VoIP Call Setup Case VoIP Call Setup Case StudiesStudies Pure SIP or H.323 Call Flow (Eyers)

Simulation based on Internet trace analysis

95th percentile of < 1 s mostly achieved for simple SIP setups

Results vary for more complex paths or over the public Internet.

Compared with TCP-based H323 call setup, retransmission timer is important.

VoIP Call Setup Case VoIP Call Setup Case StudiesStudies SIP call setup in 3GPP (Kist)

Based on traces and added DNS delay estimation

Applied in 3GPP architecture with roaming Simulation shows a mean delay of around

1800ms. SIP call setup in 3GPP rel.99. (Curcio)

Simulation shows call setup over 3GPP rel99 emulator is close to that in national Intranet calls. Delay below 100ms.

call setup over emulated low bit-rate channel (up to 2kbps) is below 1 second.

Wrap-up RemarksWrap-up Remarks Make the target of VoIP calls no worse

than those of PSTN/ISDN. E721 target values seem to be longer

than what many case studies have shown.

Allocation of delay budget depends on the VoIP call architecture and whether PSTN is involved, may refer to PSTN national/ international allocation for various VoIP network domains.

Delay Measurement Traces Traffic model and link loading for Queuing

analysis

To be discussed To be discussed ENUM query measurement

Detailed logistics to test ENUM in Fibernet infrastructure and the testing plan.

Call setup delay The architecture assumption for end-to-end

VoIP/ENUM network for call setup delay analysis.

The measurement proposal – what to measure (capacity? Delay?) and the scale (end-to-end? Components?)

Anything else that might be interested

Thank You!Thank You!