Online parameter tuning of the flow classification method in the energy-efficient data center network “HOLST”

Masaki Murakami; Hiroki Kubokawa; Kyosuke Sugiura; Eiji Oki; Satoru Okamoto; Naoaki Yamanaka

doi:10.1364/JOCN.395798

Journal of Optical Communications and Networking
Vol. 12,
Issue 11,
pp. 344-354
(2020)
•https://doi.org/10.1364/JOCN.395798

Online parameter tuning of the flow classification method in the energy-efficient data center network “HOLST”

Masaki Murakami, Hiroki Kubokawa, Kyosuke Sugiura, Eiji Oki, Satoru Okamoto, and Naoaki Yamanaka

Not Accessible

Your library or personal account may give you access

Get PDF
Email
Share
Get Citation
Copy Citation Text
Masaki Murakami, Hiroki Kubokawa, Kyosuke Sugiura, Eiji Oki, Satoru Okamoto, and Naoaki Yamanaka, "Online parameter tuning of the flow classification method in the energy-efficient data center network “HOLST”," J. Opt. Commun. Netw. 12, 344-354 (2020)

Export Citation
- BibTex
- Endnote (RIS)
- HTML
- Plain Text
Citation alert
Save article

Abstract

An online parameter-tuning method for the energy-efficient data center network (DCN) named HOLST (high-speed optical layer 1 switch system for time-slot-switching-based optical data center networks) is proposed. HOLST comprises an optical circuit switching network, optical slot switching network, and electrical packet switching network for the spine layer. It requires flows to be assigned to optimal switches to achieve high power savings. In this study, first we experimentally confirm that the change in the DCN characteristics in the short term of actual data center traffic downgrades the accuracy of flow classification. Subsequently, we propose a procedure for reconfiguring a flow classification function and a method for online parameter tuning for this function. Finally, the accuracy of the flow classification method using the proposed tuning and estimated switching energy consumption in the spine layer are evaluated. The proposed online parameter-tuning function increases the accuracy of flow classification and reduces switching energy consumption relative to the conventional flow classification function.

Full Article | PDF Article

More Like This

HiFOST: A Scalable and Low-Latency Hybrid Data Center Network Architecture Based on Flow-Controlled Fast Optical Switches

Fulong Yan, Xuwei Xue, and Nicola Calabretta
J. Opt. Commun. Netw. 10(7) B1-B14 (2018)

OPSquare: A Flat DCN Architecture Based on Flow-Controlled Optical Packet Switches

Fulong Yan, Wang Miao, Oded Raz, and Nicola Calabretta
J. Opt. Commun. Netw. 9(4) 291-303 (2017)

Virtual Network Reconfiguration for Reducing Energy Consumption in Optical Data Centers

Yuya Tarutani, Yuichi Ohsita, and Masayuki Murata
J. Opt. Commun. Netw. 6(10) 925-942 (2014)

Previous Article Next Article

Cited By

You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

Figures (20)

You do not have subscription access to this journal. Figure files are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

Tables (10)

You do not have subscription access to this journal. Article tables are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

Equations (5)

You do not have subscription access to this journal. Equations are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.

Contact your librarian or system administrator
or
Login to access Optica Member Subscription

HOLST	Number of Ports
EPS	1024
OSS	256
OCS	256
Helios
EPS	1280
OCS	256
Conventional DC
EPS	1536

Number of servers per rack	30 servers
Flow size	200 KB
Number of flows per rack	10 flows
Distribution of the destination server	Multinomial

Item	Distribution	Frontend	Backend
Inter-arrival	Exponential	$μ = 2$ [ms]	$μ = 150$ [ms]
Size	Pareto	$a = 1.5$ $μ = 50$ [KB]	$a = 1.2$ $μ = 10$ [MB]
Duration	Pareto	$a = 1.2$ $μ = 1$ [s]	$a = 1.2$ $μ = 10$ [s]
External rack	Binomial	$p = 0.95$
To Internet	Binomial	$p = 0.95$
Destination	Multinomial

Item	Distribution	Frontend	Backend
Inter-arrival	Exponential	$μ = 2$ [ms]	$μ = 150$ [ms]
Size	Pareto	$a = 1.5$ $μ = 10$ [KB]	$a = 1.2$ $μ = 100$ [MB]
Duration	Pareto	$a = 1.2$ $μ = 1$ [s]	$a = 1.2$ $μ = 10$ [s]
External rack	Binomial	$p = 0.2$
To Internet	Binomial	$p = 0.0$
Destination	Multinomial

MEMS
BOX	50 W
Element	$0.1 N$ W
$1 \times N$ PLZT [29]
Main board	2.4 W
Element + sub-board	$0.12 (N - 1)$ W
$N \times N$ PLZT [29]
Main board	2.4 W
Element + sub-board	$0.24 N (N - 1)$ W
Electrical [13]	$12.5 N$ W

1	set parameters as [MF-DF threshold, DF-EF threshold,MF queue size, DF queue size, EF queue size]
2	Require: change values, convergence threshold for each parameter
3	for parameter in parameters
4	set direction as up
5	set step as 0
6	while change value >= convergence threshold
7	Require: current correct classification rate
8	if step is 0 then
9	set previous correct classification rate as current correct classification rate
10	step ← step + 1
11	continue
12	else
13	if current correct classification rate > previous correct classification rate then
14	if direction is up then
15	value of parameter ← value of parameter $+$ change value
16	else
17	value of parameter ← value of parameter − change value
18	end if
19	else
20	change value ← change value/2
21	if direction is up then
22	set direction as down
23	value of parameter ← value of parameter −change value
24	else
25	set direction as up
26	value of parameter ← value of parameter $+$ change value
27	end if
28	end if
29	end if
30	if parameter is any queue size
31	change size of the other queues so that the total size of the three queues is constant
32	reconfigure the parameters of flow classification function
33	end while
34	end for

Traffic Model	Actual
Initial MF-DF threshold	90 packets
Initial DF-EF threshold	19,000 packets
Initial mice queue size	39 IDs
Initial doggy queue size	51 IDs
Initial elephant queue size	10 IDs
Initial change value of the MF-DF threshold	10 packets
Initial change value of the DF-EF threshold	1000 packets
Initial change values of queues	4 IDs
Convergence thresholds	Half of the initial change values
Reconfiguration period	10 s

Initial MF-DF threshold	25
Initial DF-EF threshold	9000
Initial mice queue size	33 IDs
Initial doggy queue size	53 IDs
Initial elephant queue size	14 IDs

Convergence Threshold	Time to Adapt	Correct Classification Rate
Initial change value	210 s	70.78%
Half of initial change value	300 s	71.97%
Quarter of initial change value	410 s	70.29%
1/8 of initial change value	620 s	68.11%
1/16 of initial change value	910 s	65.52%

Traffic Model	MF-DF Threshold	DF-EF Threshold	Mice Queue Size Ratio	Doggy Queue Size Ratio	Elephant Queue Size Ratio
Actual	40	11,000	0.39	0.51	0.10
Dense	25	9000	0.33	0.53	0.14

Traffic Model	MF-DF Threshold	DF-EF Threshold	Mice Queue Size Ratio	Doggy Queue Size Ratio	Elephant Queue Size Ratio
Actual	40	11,000	0.39	0.51	0.10
Dense	25	9000	0.33	0.53	0.14

Abstract

Cited By

Figures (20)

Tables (10)

Equations (5)

Journal of Optical Communications and Networking