experimental evaluation of application triggered flow...

3
OECC/PSC 2019 ©IEICE Experimental Evaluation of Application Triggered Flow Classification Using Operated Data Center Traffic Data Masaki Murakami, Masahiro Matsuno, Satoru Okamoto, and Naoaki Yamanaka Graduate School of Science and Technology, Keio University, 3-14-1 Hiyoshi, Kohoku-ku, Yokohama, Kanagawa, 223-8522 Japan [email protected] Abstract: The prototype system of the energy efficient data center network HOLST was introduced and estimated 49% reduction of switching power consumption with application triggered flow classification using operated data center traffic data. Keywords: Software defined networking (SDN) for computing and big data control INTRODUCTION With the expansion of data center (DC) use, power consumption of data center network (DCN) is rapidly increasing. Fig.1 shows leaf-spine DCN architecture. 85% of the power consumption of DC equipment is for servers and disks and the rest are for network [1]. The power consumption in Fig. 1 is assumed to be the power consumption of leaf layer and spine layer are almost equal, so the power consumption of spine switches (Spine SWs) is 8% and the power consumption of Leaf SWs is 7%. In order to reduce switching power consumption of Spine SWs, optical circuit switching (OCS) is considered to introduce to DCN because switching power consumption is much smaller than electric packet switching (EPS) [2]. DCN named HOLST (High-speed optical layer 1 switch system for timeslot switching based optical data center networks) that utilizes optical resources more efficiently by introducing optical slot switching (OSS) to the data center have been proposed [3]. The traffic of HOLST is categorized into a small capacity flow named Mice Flow (MF) accommodated in EPS, medium capacity flow named Doggy Flow (DF) accommodated in OSS and large capacity flow named Elephant Flow (EF). As a flow classification method in HOLST, a flow classification using hierarchical least recently used (LRU) queues [4] have been proposed. We have experimentally introduced the HOLST prototype system which performs flow classification using the hierarchical LRU queues at the actual data center. In this paper, we describe the experimental environment and describe the estimation result of the switching power consumption reduction effect. Fig. 1. Leaf-spine DCN architecture EXPERIMENTAL IMPLEMENTATION A. HOLST prototype system Fig. 2 shows the concept of HOLST. The spine layer of HOLST is composed of EPS, OSS, and OCS, aiming at reducing switching power consumption. The flow control of HOLST is as follows. The network accommodating the flow that reaches the Top of Rack switch (ToR SW) from the server is determined by the flow classification function implemented in the hardware of the ToR SW. The flow classification result is sent to the Software Defined Networking (SDN) controller. The SDN controller changes the configuration of ToR SW and Spine SW based on the result. The flow is accommodated in the network corresponding to the classification result and reaches the destination ToR SW. Fig. 3 shows HOLST prototype system in the actual DC. The flow classification machine classifies the flow obtained by port mirroring and sends the classification result to the Ryu SDN controller. The Ryu SDN controller configures the VLAN setting of ToR SW. The flow reaches the destination ToR SW by VLAN forwarding.

Upload: others

Post on 04-Jul-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Experimental Evaluation of Application Triggered Flow ...biblio.yamanaka.ics.keio.ac.jp/file/OECCPSC2019_murakami.pdf · OECC/PSC 2019 ©IEICE EVALUATION The obtained byte count is

OECC/PSC 2019

©IEICE

Experimental Evaluation of Application Triggered Flow Classification Using Operated

Data Center Traffic Data

Masaki Murakami, Masahiro Matsuno, Satoru Okamoto, and Naoaki Yamanaka Graduate School of Science and Technology, Keio University, 3-14-1 Hiyoshi, Kohoku-ku, Yokohama, Kanagawa, 223-8522 Japan

[email protected]

Abstract: The prototype system of the energy efficient data center network HOLST was introduced and estimated 49% reduction of switching power consumption with application triggered flow classification using operated data center traffic data. Keywords: Software defined networking (SDN) for computing and big data control

INTRODUCTION

With the expansion of data center (DC) use, power consumption of data center network (DCN) is rapidly increasing. Fig.1 shows leaf-spine DCN architecture. 85% of the power consumption of DC equipment is for servers and disks and the rest are for network [1]. The power consumption in Fig. 1 is assumed to be the power consumption of leaf layer and spine layer are almost equal, so the power consumption of spine switches (Spine SWs) is 8% and the power consumption of Leaf SWs is 7%. In order to reduce switching power consumption of Spine SWs, optical circuit switching (OCS) is considered to introduce to DCN because switching power consumption is much smaller than electric packet switching (EPS) [2]. DCN named HOLST (High-speed optical layer 1 switch system for timeslot switching based optical data center networks) that utilizes optical resources more efficiently by introducing optical slot switching (OSS) to the data center have been proposed [3]. The traffic of HOLST is categorized into a small capacity flow named Mice Flow (MF) accommodated in EPS, medium capacity flow named Doggy Flow (DF) accommodated in OSS and large capacity flow named Elephant Flow (EF). As a flow classification method in HOLST, a flow classification using hierarchical least recently used (LRU) queues [4] have been proposed. We have experimentally introduced the HOLST prototype system which performs flow classification using the hierarchical LRU queues at the actual data center. In this paper, we describe the experimental environment and describe the estimation result of the switching power consumption reduction effect.

Fig. 1. Leaf-spine DCN architecture

EXPERIMENTAL IMPLEMENTATION

A. HOLST prototype system

Fig. 2 shows the concept of HOLST. The spine layer of HOLST is composed of EPS, OSS, and OCS, aiming at reducing switching power consumption. The flow control of HOLST is as follows. The network accommodating the flow that reaches the Top of Rack switch (ToR SW) from the server is determined by the flow classification function implemented in the hardware of the ToR SW. The flow classification result is sent to the Software Defined Networking (SDN) controller. The SDN controller changes the configuration of ToR SW and Spine SW based on the result. The flow is accommodated in the network corresponding to the classification result and reaches the destination ToR SW. Fig. 3 shows HOLST prototype system in the actual DC. The flow classification machine classifies the flow obtained by port mirroring and sends the classification result to the Ryu SDN controller. The Ryu SDN controller configures the VLAN setting of ToR SW. The flow reaches the destination ToR SW by VLAN forwarding.

Page 2: Experimental Evaluation of Application Triggered Flow ...biblio.yamanaka.ics.keio.ac.jp/file/OECCPSC2019_murakami.pdf · OECC/PSC 2019 ©IEICE EVALUATION The obtained byte count is

OECC/PSC 2019

©IEICE

Fig. 2. The concept of HOLST

Fig. 3. HOLST prototype system in the actual DC

B. Flow Classification Method

Fig. 4 shows the concept of the flow classification method using Hierarchal LRU queues and Fig. 5 shows a flowchart of flow classification function in this experiment. Table 1 shows the parameters of the flow classification function in this experiment.

Fig. 4. The concept of flow classification function

Fig. 5. Flow chart of software function of flow classification machine

TABLE I PARAMETERS OF THE FLOW CLASSIFICATION FUNCTION

Parameter Value

Period of reconfiguration 5 seconds Elephant queue size 1 Flow ID Doggy queue size 4 Flow IDs Mice queue size 15 Flow IDs

Threshold of shifting from mice queue to doggy queue

60,000

Threshold of shifting From doggy queue to elephant queue

600,000

Page 3: Experimental Evaluation of Application Triggered Flow ...biblio.yamanaka.ics.keio.ac.jp/file/OECCPSC2019_murakami.pdf · OECC/PSC 2019 ©IEICE EVALUATION The obtained byte count is

OECC/PSC 2019

©IEICE

EVALUATION

The obtained byte count is shown in Fig. 5. It was confirmed that the experimental system classified flows. However, there is a peak flow to be accommodated in the optical switching in traffic flowing from electrical spine SW to ToR SW #2. This is because a flow ID should be uniquely determined from the source IP address and the destination IP address, but in this experiment, it was classified into 20 types of flow IDs according to the destination IP address, so the packet to be defined as packets defined as one flow were too coarse.

Fig. 6. Byte count of the ports of the ToR SWs

Based on [5], the estimated switching power consumption per bit is 0.365 nJ / bit for optical switches and 6.88 nJ / bit for electric switches, and Table 2 shows the results of estimating the switching power consumption of each switch. Table 3 shows the results of estimating the total switching power consumption of spine SW when all the traffic is accommodated by EPS and comparing it with the HOLST prototype system. We confirmed 49% switching power consumption reduction of Spine SWs by accommodating 5 flow IDs among 20 flow IDs determined based on the destination IP address.

TABLE 2 ESTIMATION OF SWITCHING POWER CONSUMPTION OF EACH SWITCH

Switch Switching power consumption

Electrical spine SW 113.38 kJ Optical circuit SW 5.37 kJ

Optical slot SW 1.00 kJ

TABLE 3 COMPARISON OF ESTIMATED SWITCHING POWER CONSUMPTION WITH CONVENTIONAL DC

DCN Architecture Switching power consumption

Conventional DC using only EPS 233.59 kJ HOLST 119.75 kJ

CONCLUSIONS

We have experimentally introduced the HOLST prototype system which uses flow classification method using the hierarchical LRU queues in the actual data center. In this paper, we introduced the implementation of flow classifier and experimental environment. We achieved 49% estimated value of switching power consumption reduction effect compared with DC using only conventional EPS. This contributes to a 4% reduction in the overall DCN power consumption.

ACKNOWLEDGMENT

This work is supported by “HOLST (High-speed Optical Layer 1 Switch system for Time slot switching based optical data center networks) Project” funded by New Energy and Industrial Technology Development Organization (NEDO) of Japan

REFERENCES

[1] T. Hartno, “Green Data Center,” https://www.cisco.com/c/dam/global/en_id/training-events/cnsf2008/files/Cisco_Green_Data_Centre.pdf

[2] N. Farrington, G. Porter, S. Radhakrishnan, H. H. Bazzaz, V. Subramanya, Y. Fainman, G. Papen, and A. Vahdat, “Helios: a hybrid electrical/optical switch architecture for modular data centers,” ACM SIGCOMM Computer Communication, Review, vol. 41, no. 4, pp. 339–350, 2011.

[3] M. Hirono, T. Sato, J. Matsumoto, S. Okamoto and N. Yamanaka, "HOLST: Architecture design of energy-efficient data center network based on ultra High-speed Optical Switch," 2017 IEEE International Symposium on Local and Metropolitan Area Networks (LANMAN), Osaka, 2017, pp. 1-6.

[4] Yukihiro Imakiire, Masayuki Hirono, Masaki Murakami, Satoru Okamoto, and Naoaki Yamanaka, “Flow/Application Triggered SDN control in Hybrid Data-center Network "HOLST",” 41st The Optical Fiber Communication Conference and Exhibition (OFC2018), Tu3D.6, March 2018.

[5] S. J. B. Yoo, “Energy efficiency in the future internet: The role of optical packet switching and optical-label switching,” IEEE Journal of Selected Topics in Quantum Electronics, vol. 17, no. 2, pp. 406–418, March 2011.