Transcript Document

Energy-Efficient Solutions
for 10Gbps Ethernet
Yury Audzevich, Alan Mujumdar,
Philip Watts, Andrew W. Moore
MSN 2012 workshop
Friday, July 13th, 2012
Introduction
Energy-efficiency of transmission systems is one of the key priorities with respect
to the next generation of networking equipment.
Open questions:
What is the power contribution
of the ‘lower layer’ transmission
protocols?
What is the power impact of
Servers
vs Network for a Google cluster (fat trees topology)
the encoding
blocks?
(Energy proportional datacenter networks, Abts et al. ISCA 2010)
The focus of the research
Which effect DC-balanced codes do have on the optical transmission system?
…and in particular: what is
- the effect on optical power requirement?
- front-end power contribution, like PMA and PMD?
- the power consumption of line coding block itself?
???
???
10Gb/s optical link simulations
Optical link – transmission system:
• 219 bits PRBS is used as an input, the baud rate is adjusted after encoding
Optical link parameters:
• 100m Single Mode Fibre with parameters satisfying requirements for 10Gbps
Ethernet over SMF
Optical link – receiving system:
• Optical receiver with direct detector and AC coupling achieved using High Pass Filter
• BER is calculated using the complementary error function
10Gb/s optical link simulations (cont.)
better
better
• The transmission system is relatively insensitive to the DC-balanced codec choice
• Taking 100MHz HPF cut-off frequency and assuming 20dB link budget, the laser
power requirement is lower for encoded sequences (0.3mW of savings) in
comparison to PRBS
Physical Coding Sublayer – 8B10B (and 64B66B)
8B10B line code:
1) Encoder/Decoder – implemented 3B4B and 5B6B codes plus disparity control
check for DC-balance
64B66B line code:
1) Encoder/Decoder – decoding from XGMII into 10GBASE-R format
2) Scrambler/Descrambler – mixing of data to avoid long sequences of 0s/1s
• Codecs were implemented in Verilog HDL and Synthesized using 90nm
and 45nm technical process libraries
• Industry standard estimation tools were used for power measurements
Early-days results – 8B10B PCS power
10Gb/s link results: obtained for 30 microseconds simulation periods, with a
symbol clock frequency of 625MHz for both 45nm and 90nm tech. process
Inverse of
energy-proportionality
• IDLE sequences costs MORE to encode
• Low leakage 45nm library provides decrease in power by a factor of 2
Power estimates – 64B66B PCS
10Gb/s link results: Identical pattern sent for 30 microseconds of simulation
periods, with a symbol clock frequency of 156.25 MHz
64B66B
10GBASE-SR / 10GBASE-LR
(commonly used)
8B10B
10GBASE-LX4
(less common)
• Power consumption of 64B66B codec is actually more data proportional
• Scrambling & gearbox modules have a fixed power cost (~1.5-2.5 times larger
power dissipation than the combination of encoder & decoder power)
Physical Medium Attachment – 8B10B and 64B66B
PMA components are built using both CMOS and MCML logic families
• CMOS designs were synthesized using standard cell libraries
• MCML designs were built, optimized and analysed using HSPICE tools
PMA power – 8B10B and 64B66B
64B66B
• MCML power is independent of the operating frequency but is strongly related to the
optimization criteria
• At high clock frequencies MCML designs become more power efficient than their CMOS
counterparts
• Even well power-optimized PMA designs may require 5x-10x times higher power than the
corresponding PCS blocks!
Implications
• Our recent analysis of realistic trace data(10Gbps) showed average link
utilization of only 8.79% - in concordance with [1].
• The majority of the networks are overprovisioned to sustain peak loads and
underutilized most of the time
• Current implementations of Ethernet standards require continuous
transmission of IDLE code words (even in the absence of MAC traffic)
So…
May be, we need a system that has good energy-proportionality and can
quickly restart
Sounds like we need a new MAC…
[1] T. Benson, A. Akella, and D. A. Maltz. Network traffic characteristics of data centers in the wild. In Proceedings of ACM
IMC '10, pp. 267-280, New York, USA, 2010.
Or an old energy-efficient MAC
Remember this one? Ethernet – CSMA/CD
Many features/ideas we don’t want, but one we do:
• Preambles give clocks valuable re-sync. time and allow photonic systems to
turn back on
Energy-efficient MAC
Where do we get energy-savings from:


Powering down the codecs when no data is present
Using a synchronization preamble prior to data transmission for fast CDR
Is the protocol going to make a difference?
89%
93%
93%
87%
With avg. Ethernet frame size of 1150bytes and 64bits of preamble, the effective energysaving is ~93%
Key take-aways
 Optimal laser power is independent of the DC-balanced codec chosen
 Codec power consumption is not always data-proportional
 Serialization/deserialization power dominates over all the other power
groups
 New MACs (off when idle) do save significant power
How do we test, build, trial?
regular NICs don’t help
Need something programmable but FAST…
Thank you!
QUESTIONS?