Packet Coalescing for Dual-Mode Energy Efficient Ethernet: A - - PowerPoint PPT Presentation

Packet Coalescing for Dual-Mode EEE Slide 1 of 11

Packet Coalescing for Dual-Mode Energy Efficient Ethernet: A Simulation Study

Mehrgan Mostowfi

School of Mathematical Sciences University of Northern Colorado Greeley, Colorado, USA mehrgan.mostowfi@unco.edu

Packet Coalescing for Dual-Mode EEE Slide 2 of 11

What is EEE?

An IEEE Standard (ratified in 2010) [1]

• Introduces a Low-Power Idle (LPI) mode to Ethernet links
• Works as follows for 10 Gb/s and less:

[1] IEEE P802.3az Energy Efficient Ethernet Task Force. [2] P. Reviriego, J. Hernandez, D. Larrabeiti, and J. Maestro, “Performance Evaluation of Energy Ecient Ethernet,” IEEE Communications Letters, vol. 13, pp. 697-699, September 2009.

High transition times even for 10 Gb/s and less [2]. Would be magnified for 40 Gb/s and above.

Packet Coalescing for Dual-Mode EEE Slide 3 of 11

EEE for 40 Gb/s and above

Two modes are introduced [1]

• Deep Sleep: Identical to LPI
• Fast Wake:
• Does not yield any power savings
• In sending side to keep the sender and receiver in alignment
• Much faster transitions (Wakeup: 0.34 us vs. 5.50 us from DS)
• Limits power savings to receive side only in practice

[1] IEEE 802.3bj-2014 Amendment 2: Physical Layer Specications and Management Parameters for 100 Gb/s Operation Over Backplanes and Copper Cables,“ IEEE Computer Society, June 2014. [2] H. Barrass, “Options for EEE in 100G,” presentation at IEEE P802.3bj meeting, January 2012.

Can Fast Wake effectively become a new low-power mode for EEE? A case can be made [2]

• Only needs to send LPI signals

Transition times would still be too high.

Packet Coalescing for Dual-Mode EEE Slide 4 of 11

Overcoming High Transition Times

Packet Coalescing

• Gather up a few packets, send in one burst
• A sleep and wakeup transition for a number of packets
• Time-based and Count-based
• Has shown to be effective for EEE for 10 Gb/s and below [1]

[1] K. Christensen, P. Reviriego, B. Nordman, M. Bennett, M. Mostowfi, and J. Maestro, “IEEE 802.3az: The Road to Energy Efficient Ethernet,” IEEE Communications Magazine, vol. 48, no. 11, pp. 50-56, November 2010.

Power Active (100%) Fast Wake (70%) Deep Sleep (10%)

Coalescing begins Fast Wake Timer expires, pkts in buffer < Scoal/2 Packet transmission Packets arrive, coalesced Packet transmission Packets arrive, coalesced Coalescing timer expires, pkts in buffer ≥ Scoal/2 Coalescing timer expires, pkts in buffer < Scoal/2 Fast Wake Timer expires, pkts in buffer ≥ Scoal/2 Packets arrive, coalesced

TAtoF TFtoA TAtoD TDtoA TAtoD TDtoA TAtoF TFtoA TAtoF TFtoD Tcoal Tcoal Tidle Tidle

Coalescing begins

Tidle

Coalescing begins

Packet Coalescing for Dual-Mode EEE Slide 5 of 11

Simulation Model

CSIM simulation library in C

• A unidirectional EEE link with Active, FastWake, and Deep

Sleep modes

• Ethernet link was simulated by a CSIM server facility
• A Finite State Machine (FSM) controlled the server
• Some variables and timers in the FSM:

Packet Coalescing for Dual-Mode EEE Slide 6 of 11

The Finite State Machine

(Buffer empty) && (DFlag == FALSE) 3) ------------------------- Set TTIMER to TAtoF Start TTIMER

Deep Sleep Fast Wake toFW

TTIMER expired 6) ------------------------- Set FTIMER to Tidle Start FTIMER (FTIMER expired) && (Buffer is empty) 9) --------------------------- Set TTIMER to TFtoD Start TTIMER TTIMER expired 16) --------------------- (CTimer Expired) || ((# of packets in buffer ≥ Scoal) 14) ----------------------------------------------------------------- DFLag = (# of packets in buffer < (Scoal/2)) Set TTIMER to TDtoA Start TTIMER (FTIMER expired) && (Buffer is not empty) 8) ------------------- DFLag = (# of packets in buffer < (Scoal/2)) Set TTIMER to TFtoA Start TTIMER TTIMER expired 11) -------------------------

Active

Packet arrived 1) ------------------- Buffer packet Packet in buffer 2) --------------------- Transmit packet

RESET

Packet arrived 5) -------------------- Buffer packet

toDS toA

Packet arrived 15) -------------------- Buffer packet Packet arrived 10) -------------------- Buffer packet Packet arrived 7) -------------------- Buffer packet (Packet arrived) && (Buffer is not empty) 13) ---------------------------- Buffer packet (Buffer empty) && (DFlag == TRUE) 4) ------------------------- Set TTIMER to TAtoD Start TTIMER (Packet arrived) && (Buffer is empty) 12) -------------------- Buffer packet Set CTIMER to Tcoal Start CTIMER

Packet Coalescing for Dual-Mode EEE Slide 7 of 11

Experiments

Performed on the Simulation Model

• Link capacity: 40 Gb/s
• Transition times: taken from the standard (closest in meaning)
• Coalescing time Fast Wake: 3.00 us
• Power consumptions:
• Active and transitions: 100% (the peak power consumption)
• Fast Wake: 70%
• Deep Sleep: 10%
• Two set of experiments:
• Smooth Traffic: Poisson packet arrivals.
• Bursty Traffic: Interrupted Poisson Process packet arrivals
• Two coalescer sizes:
• Small: In Deep Sleep coalesce 10 pkts or for 3.00 us
• Large: In Deep Sleep coalesce 100 pkts or for 30.00 us

Packet Coalescing for Dual-Mode EEE Slide 8 of 11

Results – Poisson Traffic

Coalescing brings the power consumption closer to ideal, compared to Deep Sleep Only and no coalescing Tradeoff: added per-packet delay

Packet Coalescing for Dual-Mode EEE Slide 9 of 11

Results – Bursty Traffic

Similar results. Better consumption is the traffic is bursty (some of the coalescing is done already) Is the added delay significant?

Packet Coalescing for Dual-Mode EEE Slide 10 of 11

Conclusions and Future Work

Conclusions:

• Dual-Mode EEE can be an effective method of

extending idle periods between packet arrivals in order to maximize the opportunity for sleeping.

• The tradeoff, the increased packet delay, may be

justifiable by the energy savings gained. Future Work:

• The technical feasibility? (more EE than CS)
• Other response variables?
• Other tradeoffs?
• Real traffic traces?

Packet Coalescing for Dual-Mode EEE Slide 11 of 11

Questions from the audience

Thanks for listening. Any questions?

Mehrgan Mostowfi

School of Mathematical Sciences University of Northern Colorado Greeley, Colorado, USA mehrgan.mostowfi@unco.edu See the paper for more details