commercial applications of multi-core at ericsson

Slide titleIn CAPITALS

50 pt

Slide subtitle 32 pt

Commercial Applications of Multi-core at

EricssonHans Nilsson,

[email protected]

Top right corner for field-mark, customer or partner logotypes. See Best practice for example.

Slide title 40 pt


Text 24 pt

Bullets level 2-520 pt

Sixth ACM SIGPLAN Erlang Workshop (c) Ericsson 2007 Commercial Applications of Multi-core at Ericsson October 5, 20072

Ericsson

1876 Customers: mobile and fixed network operators Over 1,000 networks in 140 countries 40 percent of all mobile calls are made through our

systems 65507 emplyees 50% of


Slide title 40 pt


Text 24 pt



Mobile Phones are no fun anymore

The original mobile phone, 1956


Slide title 40 pt


Text 24 pt



A new communications architecture

Tra

nsp

ort

net

wo

rk

Multi Access EdgeMulti Access Edge

IMS and communication enablersIMS and communication enablers

Service networkMultimedia

Network

Users

Fixed accessFixed access Mobile accessMobile access

Softswitch and IMS, first steps towards an all-IP network


Slide title 40 pt


Text 24 pt



Modern Telecom

Lots of old phones

IP-phones

Communication clients on PCs

3G Phones

Telecom network

?

SIP, RTPIS

UP

, PC

M

GSM

UMTS

SIP

, RT

P

SIP

, R

TP

SIP, RTP

H.3

23, R

TP


Slide title 40 pt


Text 24 pt



Modern Telecom (cont’d)

Session handling

Services

Media processing RTP

RTP

SIP

SIP

SIP

H.2

48

Diameter

Many; Variationse.g. Java-based

frameworks

Complex FSMsHigh degree of concurrency

Wire-speedNetwork processors,

DSPs, ...

Billing

ISUP

Huge data volumes


Slide title 40 pt


Text 24 pt



Different Execution Platforms

... Highly scalable processingclusters (C++/Erlang/3PP)

Compact ”system on a blade”Tight space and power budget

http://www.ericsson.com/ericsson/corpinfo/publications/review/2005_01/files/2005014.pdf

Subracks of dedicated blades(e.g. ”Integrated Site”, or IS)

CPU

DP

NPDSPs


Slide title 40 pt


Text 24 pt



CPU

DP

NPDSPs






Multicore can help reduce footprint• E.g. through virtualization• Software already written for

distributed processing


Slide title 40 pt


Text 24 pt








Multicore can help through• Packing more umph in one slot• Replacing different special-purpose

processors, reducing cost and footprint

CPU

DP

NPDSPs


Slide title 40 pt


Text 24 pt



CPU

DP

NPDSPs






Multicore can help through• Making each ”blade system” more

compact• Better performance/slot

=> better price/performance• (Chaining subracks is always tricky

– inter-subrack links a bottleneck)


Slide title 40 pt


Text 24 pt



Difficult(?)Difficult(?)

Different levels of parallelism

OS OS OSVirtualContainer:

App App AppOperatingSystem:

Thread Thread ThreadApplication:

Code/DataSegm

Code/DataSegm

Code/DataSegmThread:


Slide title 40 pt


Text 24 pt



Many (≥105) independent threads Few supervisors, many workers (instances of the same

code)

No matrix operations No long loops with inherenet parallellism

Our parallellism


Slide title 40 pt


Text 24 pt



Erlang program features

Erlang already modelled ”with annotations” in form of processes

Processes models real world activities– C(++) is not like that


Slide title 40 pt


Text 24 pt



Easy!Easy!

Difficult(?)Difficult(?)

Different levels of parallelism

OS OS OSVirtualContainer:

App App AppOperatingSystem:

Thread Thread ThreadApplication:

Code/DataSegm

Code/DataSegm

Code/DataSegmThread:


Slide title 40 pt


Text 24 pt



What is left to do then?

Analysis of dataflow //– pattern matching….

Bandwidth? Memory bandwidth?


Slide title 40 pt


Text 24 pt



Case study: The Ericsson TGC


Slide title 40 pt


Text 24 pt



The TGC

Telephony Gateway Controller Formerly known as

– the Hybrid reported (by customer!) to have ISP of 0.999999999

– the Mediation Logic

Runs on AXD 301 and IS Developed in Erlang Multicore version shipped to customer

AXETGC

GWGW GW


Slide title 40 pt


Text 24 pt



The TGC (internal)

Process model– Servers– Workers

~100 servers ~10 workers per call 10,000 workers is

perfectly normal

AXE

GW GW

dispatcher

c-link half-call

p-link

half-callhalf-callhalf-call

half-call

half-call

decoderdecoderdecoderp-link


Slide title 40 pt


Text 24 pt



Erlang on multicore

Erlang programs are meant for distributed processing

– Erlang processes do not share data

– Message passing is distribution transparent

SMP prototype -97, serious implementation in -05.

Mid -06 we ran a benchmark mimicking call handling (axdmark) with a prototype SMP emulator. Observed speedup/core: 0.95

TGC released on multicore in Q207

Tim

e (s

)

Simultaneous processes16 schedulers

http://www.franklinmint.fm/blog/archives/000792.html

”Big bang” benchmark on Sunfire T2000

1 scheduler


Slide title 40 pt


Text 24 pt



TGC Results (top)

Tasks: 50 total, 2 running, 48 sleeping, 0 stopped, 0 zombie

Cpu0 : 62.5% us, 3.7% sy, 0.0% ni, 32.4% id, 0.0% wa, 0.0% hi, 1.3% siCpu1 : 36.1% us, 2.7% sy, 0.0% ni, 60.9% id, 0.0% wa, 0.0% hi, 0.3% siMem: 4092764k total, 459352k used, 3633412k free, 8196k buffersSwap: 0k total, 0k used, 0k free, 215796k cached

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 1975 homer 25 0 2295m 192m 2144 S 99.9 4.8 179:40.46 beam.smp 1 root 16 0 664 244 208 S 0.0 0.0 0:01.50 init 2 root RT 0 0 0 0 S 0.0 0.0 0:00.02 migration/0 3 root 34 19 0 0 0 S 0.0 0.0 0:00.00 ksoftirqd/0 4 root RT 0 0 0 0 S 0.0 0.0 0:00.01 migration/1 5 root 34 19 0 0 0 S 0.0 0.0 0:00.00 ksoftirqd/1


Slide title 40 pt


Text 24 pt



TGC Results (dtop)

ppb1_bs13-R3A@blade_ size 2345(131M, cpu% 107, procs 10371, runq 0 15:15:53

memory[kB]: proc 58223, atom 1768, bin 170, code 29772, ets 39215

pid name current msgq mem cpu<0.5872.491 prfTarg (prfPrc:pinf/2) 0 2036 22<0.18323.47 (erlang:apply/2) (gcpServ:recv1/3) 0 17 10<0.18436.47 (erlang:apply/2) (gcpServ:recv1/3) 0 24 5<0.1813.0> sysProc (gen_server:loop/6) 0 981 2<0.27384.47 (pthTcpNetHandler:init/1) (gen_server:loop/6) 0 587 1<0.18350.47 (erlang:apply/2) (gcpTransportProxy: 0 8 1<0.1935.0> ccpcServer_n (gen_server:loop/6) 0 587 0<0.18526.47 (erlang:apply/2) (gcpTransportProxy: 0 6 0<0.1923.0> sbm (gen_server:loop/6) 0 1719 0<0.3603.0> (erlang:apply/2) (gcpServ:recv1/3) 0 5 0


Slide title 40 pt


Text 24 pt



TGC results (performance)

3.17X call/sec

1.55X call/sec

0.4X call/sec

AXDCPB5

14X call/sec

7.6X call/sec

2.1X call/sec

AXDCPB6

ISUP-ISUP /Intra MGW

ISUP-ISUP /Inter MGW

POTS-POTS /AGW

Trafficscenario

5.5X call/sec

3.6X call/sec

X call/sec

IS/GCP1slot/board

7.7X call/sec

One core used

2.3X call/sec

One core used

IS/GEPDual coreOne core running

2slots/board

26X call/sec

13X call/secOTP R11_3

beta+patches

4.3X call/secOTP R11_3

beta+patches

IS/GEPDual coreTwo cores

running2slots/board


Slide title 40 pt


Text 24 pt



TGC Experience

Porting effort: negligible (for the application) Porting effort: modest (for the platform) Architecture dependency: low (for the application) Results: excellent Future: bright

“Funky languages” (Hagersten) can sometimes save the day


Slide title 40 pt


Text 24 pt




Slide title 40 pt


Text 24 pt



Conclusion

Still need much more research on low-level parallellism like pattern matching

Commercial multi-core applications is not the future.– It is NOW (iff you are an Erlang adict)


Slide title 40 pt


Text 24 pt



Questions?


Slide title 40 pt


Text 24 pt



commercial applications of multi-core at ericsson

Documents