Green computing in IEEE 802.3az enabled clusters Dimitar Pavlov Joris Soeurt SNE July 5, 2012 Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 1 / 26
Green Computing ”Data centres emit over 150 metric tons of CO2 per year, and the volume is increasing.” 1 ”Carbon dioxide emissions account for 80% of the contribution to global warming...” 2 Different strategies towards environmentally sustainable IT ◮ Computational efficiency (e.g. optimizing of algorithms) ◮ Consolidation (e.g. virtualization) ◮ Reducing / recycling of e-waste ◮ Resource allocation (e.g. route data to most green datacenter) ◮ Green networking 1 Baroudi et al. (2009) 2 Lashof et al. (1990) Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 2 / 26
Approaches to Green Networking ”Recent studies have shown that network devices comprise more than 15% of the total energy consumption of a datacenter.” 3 Adjust transmission power based on cable length ◮ Cables of 5m do not need same transmission power as 100m cables Power down circuitry when the line protocol is down ◮ If the line protocol is down, why keep the hardware active? Use signalling to put circuitry in lower power mode when idle ◮ Done by IEEE 802.3az ◮ Signalling protocol to put circuitry (of both sides of the connection) in sleep mode when the transmit buffer is empty ◮ State transitions operates on the microsecond level, and is therefore invisible to higher layers ◮ Both sides should announce 802.3az support during autonegotiation 3 Barroso et al. 2007 Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 3 / 26
Research motivation Research motivation How can an application optimize its energy effiency using the 802.3az protocol in cluster environments? How does the protocol achieve its energy savings? How to model the energy characteristics of 802.3az compliant devices? How to apply the energy model in cluster computing? Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 4 / 26
Cluster Overview W W W W W W W W W W W W Figure: Simplified model of cluster environment Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 5 / 26
Theoretical Study on 802.3az What is the background of this protocol? How does the protocol achieve its energy savings? What earlier research has been done on this protocol? Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 6 / 26
Timeline of 802.3az 8-2007 9-2008 7-2009 3-2010 11-2010 Project authorization request Draft 1.0 Draft 2.0 Draft 3.0 IEEE standard 12-2006 8-2008 Call for interest Task Force Review Researched mostly theoretically before, hardware implementations of final standard only now arriving on market 802.3az has not been researched in the context of cluster environments Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 7 / 26
802.3az link states Refresh interval ACTIVE SLEEP WAKE ACTIVE REFRESH 182 μ s 20 000 μ s 198 μ s 16.5 μ s T d T s T q T r T w power saving occurs T s = time to sleep T w = time to wake ◮ Send LPI ◮ Equals time needed for T q = time quiescent sending max size frame T r = time to refresh ◮ Detect link failure T d = decision time Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 8 / 26
802.3az projected energy savings 100 90 Energy consumption (% of peak) 80 70 60 50 40 30 No EEE 20 EEE (simulation) Ideal 10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Load Figure: Simulated energy consumption [Reviriego 2009] Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 9 / 26
Experimental Phase Observe the energy behaviour of 802.3az in different situations ◮ (devices, linkspeeds, throughput and protocols) Construct energy profiles for different devices Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 10 / 26
Equipment Overview (a) Cisco SG300-28 (b) Huawei S1728GWR-4P (c) Extreme x440-24p (d) Racktivity PDU (e) Schleifenbauer PDU (f) Realtek 8111E (g) Intel I350-T2 NICs (h) Servers Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 11 / 26
Test setup PDU ethernet 1: 18W 5: 0W 2: 0W 6:0W 3: 0W 7: 0W power 4: 0W 8: 0W VLAN VLAN Switch VLAN VLAN VLAN 4 6 8 10 2 1 3 5 7 9 11 13 15 17 19 21 23 25 27 V V V 2 4 6 8 10 12 14 16 18 20 22 24 26 28 VLAN VLAN VLAN VLAN VLAN VLAN 1 3 5 7 9 11 Controller Source Destination Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 12 / 26
Test setup Figure: What it actually looked like... Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 13 / 26
Experiments – Maximum energy savings Goal: determine maximum energy savings with 802.3az ◮ Fully utilize the switch to measure maximum consumption ◮ Measure the minimal switch consumption when no traffic is present ◮ Compare both measurements to determine maximum energy savings Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 14 / 26
Experiments – Maximum energy savings (cont’d) 21 20 19 18 Energy consumption (W) 17 16 15 14 13 12 11 10 No EEE 9 EEE 8 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 Time (Sec) Figure: Cisco SG300-28 using Iperf – TCP/UDP traffic at 1Gbps Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 15 / 26
Experiments – Throughput vs energy consumption Goal: determine the relationship between throughput and energy usage ◮ Generation of traffic is done with Iperf ◮ The transmission rate is set with tc per test run ◮ Energy usage of the switch is measured per test run ◮ Traffic is generated for 5 minutes then the measurements are averaged per run Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 16 / 26
Experiments – Throughput vs energy consumption (cont’d) Results from TCP tests only. UDP shows unexpected results. 21 20 19 Energy consumption (W) 18 17 16 15 14 13 12 11 No EEE 10 EEE 9 50 100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000 Throughput (MBit) � 10 . 9 + Thrghpt ∗ 18 . 9 − 10 . 9 if Thrghpt < 400 , 400 Figure: Cisco SG300-28, E s = 18 . 9 , if E s ≥ 400 Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 17 / 26
Experiments Summary Experiment shows that 802.3az has the potential to save power Vendor claims of 30% savings are generally true Odd power usage distribution – does not conform to previous research 4 Constructed power profiles, which were used as input to final phase 4 Riveriego, 2009 Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 18 / 26
Applying 802.3az in applications Distributed computing & Cluster computing Model for optimizing energy usage ◮ Define a way to determine energy usage with 802.3az ◮ Estimate switches energy consumption based on number of active ports ◮ Estimate time distribution for particular tasks with a focus on parallel computing ◮ Determine best transmission rate for a fixed quantity of data Combine output of all phases and create a prototype power calculator Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 19 / 26
Estimating time distribution in parallel computing 1 0.9 ask (minutes) 0.8 0.7 0.6 Time Needed for T 0.5 0.4 0.3 0.2 Time 0.1 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 Number of nodes Figure: Time Needed for Task, T t = T s n + ( n − 1) ∗ C Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 20 / 26
Total power consumption for a parallelized task ask (Switch) (Wh) 0.22 0.2 0.18 0.16 otal Energy Consumed for T 0.14 0.12 0.1 0.08 0.06 Swith(es) Energy Consumption 0.04 T 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 Number of nodes Figure: Huawei S1728GWR-4P, E task = ( P b ∗ N s + P pp ∗ N n ) ∗ T Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 21 / 26
Estimating optimal bandwidth for a fixed-datasize task ask (Wh) T otal Switch Energy Consumed for T ask 0.2 0.18 otal Switch Energy Consumed for T 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950 1000 T Speed (mbps) T otal Switch Energy Consumed Figure: Huawei S1728GWR-4P, E task = P total ∗ ( A d ∗ 8) S t Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 22 / 26
Prototype power calculator operation input output Number of nodes Power as function of Switch power profile number of nodes Ports per switch Total time for task Calculator Communication overhead Speed Power as function of Data size transmission rate Available time Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 23 / 26
Conclusions The 802.3az protocol can potentially optimize the energy efficiency of networked environments The technology, when applied within a distributed computing environment, contributes to green IT efforts 802.3az can save energy with the vast majority of traffic patterns To achieve optimal energy savings, one needs to perform low-level software changes Dimitar Pavlov, Joris Soeurt (SNE) Green computing July 5, 2012 24 / 26
Recommend
More recommend
