July 2004

SoC Architecture Simplifies Design of Integrated Voice and Data Systems For the Network Edge

BY JONATHAN PEACE

The much heralded arrival of VoIP to the enterprise brought with it the promise of large reductions in operating expense through the unification of voice and data systems, as well as the possibility of new applications previously either difficult or impossible to implement. One of these is the so called �office in a box,� a union of the traditional PBX with a WAN router, firewall, VPN, and VoIP gateway.

Exciting new applications, however, cannot sacrifice any of the performance or features which customers have come to expect from their legacy systems. In order to meet these demands, system developers need a cost-effective, flexible platform on which to build the broad variety of applications and features that users require, but which at the same time meets the highest quality of service (QoS) expectations of their customers.
 

VoIP PROCESSOR EVOLUTION
Early enterprise media gateway designs generally adopted a network processor + discrete DSP architecture. This �bolt-on� approach was suitable for low channel counts. It gradually gave way to integrated DSP and packet processors due to cost, performance, and scalability requirements. Such devices are commonly referred to as Media Stream Processors, or MSPs, and typically perform all processing of the voice path, such as echo cancellation, voice compression, tone detection, and packet loss concealment. MSPs are controlled by a separate host controller, which itself can be relieved of all fast path processing responsibility.
 

The MSP, often with integrated Ethernet or other packet interface, solves the problem of performance and scalability, but provides only part of the functionality required for the �office in a box.� The other part, the host controller, actually provides a rich set of functionality including system control, telephony, packet and router signaling, authentication, key negotiation, storage access and much more.
 

For the first time, we are starting to see the introduction of devices that integrate all of the above functions, providing a true single-chip solution for �office in a box� products.

APPLICATIONS CHALLENGES
The applications encompassed by the �office-in-a-box� processor architecture are extremely broad, covering a wide range of media processing and networking applications.

The challenge facing mixed-media equipment system designers for this broad range of applications is to provide an extremely rich variety of application-level functionality, while at the same time guaranteeing interactive media performance and system security for large numbers of simultaneous users.

OFFICE-IN-A-BOX SoC SOLUTION
The �office in a box� system-on-a chip (SoC) architecture described here solves this problem by logically and physically separating the hard real-time and soft real-time functions on the die, but provides a clean and well defined mechanism for communication between the two halves. The core of the architecture (Figure 1) is comprised of the Media Stream Processor (MSP) and the Control and Signaling Processor (CSP).

The MSP is responsible for all delay sensitive media processing within the device, and in this case contains three main elements:

� a high-performance voiceband signal processor;
� a RISC network processor with hardware acceleration for specific functions; and
� a hardware cipher engine with Internet key exchange (IKE) acceleration support.
 

The CSP is responsible for the both the control of the MSP and processing of delay insensitive, or bulk, data. In order to support this requirement it has a large memory space and MMU, and a full complement of peripherals for system expansion and control.

The two main subsystems communicate over a shared memory interface, which is abstracted in the CSP as an Ethernet interface by a virtual Ethernet driver.
The separation of the two main functions of the system has many advantages for the systems designer. The MSP includes pre-tested software that performs all of the voice and data encapsulation processing with low-latency and guaranteed performance. This binary, or microcode, performs all voice processing including echo cancellation, voice coding, and packetization, from TDM to cell or packet interface. Since the code is pre-tested and runs with guaranteed resources, it can maintain a high quality level whatever applications may be running on the CSP.

The CSP can run any common embedded operating system such as VxWorks and Linux. Applications executing on the CSP are completely independent of the MSP processing. One significant advantage of running non real-time operating systems such as Linux or OpenBSD is that it gives designers access to an enormous wealth of open-source code that can be quickly and easily ported to the CSP.

SOLVING SYSTEM DESIGN CHALLENGES
The debate over big versus little endian code is not likely to go away any time soon, and can present a problem any time code written for one endianess is mixed with code of the opposite flavor. The MSP/CSP approach elegantly solves this issue. The MSP communicates directly with all fast path peripherals but exposes all control and data planes over virtual Ethernet interfaces, limiting the inter-processor interaction to a single device driver, which can be supplied as source code with the device.

Another challenge facing designers is the debate over using a Memory Management Unit, or MMU. For feature-rich environments such as the �office in a box,� access to large program and data memory space is essential, as is the ability to run applications not permanently resident in memory, such as Java for example. For these reasons, an MMU is highly desirable. However, for the highest performance, especially as it relates to interrupt handling of small packets typical in a packet voice system, the time associated with switching MMU context can extract a significant performance penalty. The MSP/CSP architecture sidesteps this issue, by allowing each processor to do what it does best, with or without MMU.
 

The clean abstraction of the virtual Ethernet interface provides another valuable side-benefit. Existing application code can be quickly hooked up to the MSP, often achieving �first voice� in a matter of days, allowing quick and easy evaluation and prototyping of new systems.
 

SECURE WATERTIGHT COMPARTMENTS
The MSP/CSP model also provides physical separation of media stream from applications, greatly reducing vulnerability to attacks that attempt to manipulate the host processor program counter, such as �buffer overflow exploits.� This provides the system designer with an extra level of built-in security, which is especially important as more and more valuable content is streamed through the device, for example in video-on-demand applications.
 

UNIFIED MEMORY ARCHITECTURE
To reduce overall system cost, the processor design can employ a unified memory architecture in which the MSP and CSP share a single external commodity SDRAM, either SDR or DDR. In addition, a large on-chip SRAM could provide ample buffer space for the peripheral DMA engines to buffer traffic. Each of these memory subsystems should be accessed at the same time without contention.
 

PROGRAMMING MODEL
The virtual Ethernet driver source code is supplied for a number of operating systems, and can be easily ported to others. After system boot, the MSP appears to the CSP as a host on an Ethernet network. Referring to Figure 2, assume that the CSP wishes to instruct the MSP to perform the following functions:
 

� take cells from Utopia and perform AAL5;
� then look for specific RTP stream from the resulting IP packets;
� then render those packets through a G.723 codec;
� perform G.168 Echo Cancellation; and
� output voice channel to TDM timeslot 38.
 

In this case, the CSP application simply sends a set of high-level commands over one of the virtual Ethernet interfaces, with the above commands encoded using a straightforward API.
 

The virtual Ethernet model, while removing the need to write device drivers, has another side benefit: scalability. MSP processing can be scaled smoothly and with no code changes by adding extra MSP capacity externally to the device, but connected to the system master over Ethernet using an off-the-shelf Ethernet switch chip.
 

The virtual Ethernet abstraction is used not only for control but also when the CSP wishes to inspect or modify the fast-path traffic. To facilitate this, the system employs a zero-copy architecture for any fast path traffic that needs to be handled by the CSP. For instance, in the example above, after segmentation and reassembly, any of the IP traffic not destined for TDM would be passed to the CSP for further processing, e.g., firewalling, routing or network address translation (NAT). The supplied virtual Ethernet driver performs this function with no additional copies, reducing both memory bandwidth and processor overhead.
 

CONCLUSION
Integrating separate media stream and signal processors on a single chip offers a new level of silicon integration and cost effectiveness to designers of access and network edge equipment and applications. The processor architecture offers a cost effective, high-performance, and scalable platform on which to build the extremely broad variety of applications and features, which users demand, along with the carrier class voice quality and performance that those same users have come to expect.

Jonathan Peace is CTO, Mindspeed Technologies� Multiservice Access Division. Mindspeed Technologies designs, develops, and sells semiconductor networking solutions for communications applications in enterprise, access, metropolitan, and wide-area networks.
 

If you are interested in purchasing reprints of this article (in either print or HTML format), please visit Reprint Management Services online at www.reprintbuyer.com or contact a representative via e-mail at [email protected] or by phone at 800-290-5460.

[ Return To The July 2004 Table Of Contents ]