|
An emerging trend in consumer networking is to merge the broadband
modem with home networking and possibly voice support, creating
a residential gateway. Responding to this trend, semiconductor
vendors are offering processors that combine some or all of
these functions, reducing the cost of full-featured residential
gateways. Examples include Broadcom’s BCM6358, Infineon’s
Danube, PMC-Sierra’s MSP7100, Ubicom’s SE5000,
Freescale’s PowerQuicc MPC8323E, and the Puma 4 from
Texas Instruments. We call these new devices gateway processors.
Faster,
Multithreaded CPUS
VDSL,
PON, and other high-speed broadband technologies require much
more packet processing than traditional broadband. The simplest
packet-processing architecture is a single CPU that handles all
control- and data-plane tasks in software. A commodity RISC CPU
(e.g., 200MHz MIPS 4K) can route about 25,000 packets per second,
which translates to 100Mbps for average packets. When performing
tasks that require more instructions per packet, such as firewall/NAT,
the same CPU may achieve only 8,000 packets per second, cutting
the performance to 30Mbps on average packets. Adding functions
such as IPSec, ATM bonding, or quality of service (QoS) will
further reduce throughput.
Thus, for residential gateways operating at ADSL speed or 802.11b
WLANs, a single commodity CPU is adequate. But for VDSL and PON,
a commodity CPU starts to become a bottleneck. The popular MIPS
4K CPU, therefore, is being supplanted by faster CPUs. Some vendors
are licensing the faster MIPS 24K or 34K, but Broadcom, Freescale,
Intel, Marvell, and Ubicom design their own CPUs.
To achieve greater utilization, companies are starting to use
multithreading, which allows a single physical CPU to provide
two or more threads,
each of which appears to software as a separate processor.
The CPU switches among the threads every cycle; if one
thread is
stalled due to a memory access or other slow operation, the
CPU simply
executes the other threads instead. This method improves CPU
efficiency when separate tasks can be assigned to each thread.
PMC, for example,
assigns control-plane tasks to one thread, voice to another,
and data-plane tasks to other threads. Broadcom and Ubicom
also have
multithreaded processors. Scaling
Performance
Even
with multithreading, increasing CPU performance to handle faster
data rates and more services would result in an expensive chip
that dissipates too much power. It is more efficient, instead,
to dedicate specialized resources for data-plane processing.
The first step gateway-processor suppliers have taken has been
to implement Layer 2 protocol processing functions in hardwired
circuitry. DSL processors, for example, include hardware for ATM
segmentation and reassembly (SAR) and QoS to unburden the CPU.
Some of the more recent implementations also had hardware to accelerate
ATM bonding. Ethernet MACs are also implemented in hardware.
The second step has been to use programmable packet engines,
principally for Layer 3 and higher functions. Packet engines
are optimized
for net-working tasks and therefore use less die area and less
power than a comparable CPU does. A packet engine is a vital
complement to both CPUs and hardwired logic for gateways
connected to networks
at 100Mbps and faster.
Because PON gateways have the highest data rates and deliver
the most services, we find packet engines in PON controllers.
They
are also in the newest VDSL gateway processors. Dedicated hardware
should yield better performance under combined heavy loads
from control-plane, data-plane, and voice processing. VoIP
Support
Most
gateway processors announced in the past year handle voice coding.
How they implement the function, however, is changing. Earlier
devices usually used a DSP, either integrated with the voice
processor or attached externally. Now, most vendors use a CPU
instead.
Using a CPU for voice coding eliminates the die area required
for a separate voice DSP. Developing voice software on
a CPU is easier
than on a DSP. Another advantage is that CPUs support unlimited
code space through caching, whereas most DSPs have a fixed-size
code store. As the number of voice codecs increases, this code
store is often too small to hold them all.
Recent gateway processors illustrate different approaches to
CPU-based voice processing. Danube dedicates an entire CPU
core to the task,
reserving the chip’s other CPU for other tasks. Freescale
simply runs the voice software on the main CPU, leaving it up to
the operating system to ensure real-time response. Broadcom, PMC,
and Ubicom dedicate a single thread to voice processing.
CPU-based voice processing reduces the cost of VoIP support.
As a result, more gateway vendors are including VoIP in their
newest
products. For example, 36% of cable modems shipped in 1Q06
included voice support, up from 21% in 1Q05. Similarly, Infonetics
reported
42% quarter-over-quarter growth for VoIP-enabled DSL modems
in 4Q05.
Gateway processors fill an important market need. The capability
of these chips will evolve rapidly to meet the changing
requirements of broadband and home networks.
Originally published in Nikkei
Electronics Asia, July
2006
© 2002-2006 The Linley Group
|