Optical network topology design to execute many&#x00A0;tasks simultaneously in a disaggregated data&#x00A0;center

Akishige Ikoma; Yuichi Ohsita; Masayuki Murata

doi:10.1364/JOCN.524628

Journal of Optical Communications and Networking
Vol. 16,
Issue 7,
pp. 764-780
(2024)
•https://doi.org/10.1364/JOCN.524628

Optical network topology design to execute many tasks simultaneously in a disaggregated data center

Akishige Ikoma, Yuichi Ohsita, and Masayuki Murata

Not Accessible

Your library or personal account may give you access

Get PDF
Email
Share
Get Citation
Copy Citation Text
Akishige Ikoma, Yuichi Ohsita, and Masayuki Murata, "Optical network topology design to execute many tasks simultaneously in a disaggregated data center," J. Opt. Commun. Netw. 16, 764-780 (2024)

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

Check for updates

Abstract

In a disaggregated data center (DDC), task execution is reliant on the communication between resources, making performance highly sensitive to network quality. An optimized physical network topology is crucial for a DDC. To enable the simultaneous execution of numerous tasks, a substantial number of communicable resource pairs satisfying performance requirements is necessary. We propose a physical topology evaluation metric called the capability of simultaneous task execution (CSTE) and a corresponding physical topology design leveraging CSTE for a DDC equipped with optical networks. CSTE represents the ratio of resources that could be used as a resource communicating with other resources without violating the performance requirements in a situation where tasks up to the maximum number of executable tasks are executed. In addition, we formulated a physical topology design problem aimed at generating a physical network topology capable of maximizing task execution based on CSTE. By solving this optimization problem, we generated topologies and validated their effectiveness via task allocation simulations. The results showed that an optimal topology based on CSTE reduces task blockages by over 50% compared to conventional topologies. In addition, the results exhibited a positive correlation with the number of executable tasks. Through a physical topology design based on CSTE, we could construct a DDC that could handle a larger volume of tasks.

Full Article | PDF Article

More Like This

Optically Disaggregated Data Centers With Minimal Remote Memory Latency: Technologies, Architectures, and Resource Allocation [Invited]

Georgios Zervas, Hui Yuan, Arsalan Saljoghei, Qianqiao Chen, and Vaibhawa Mishra
J. Opt. Commun. Netw. 10(2) A270-A285 (2018)

DACON: a reconfigurable application-centric optical network for disaggregated data center infrastructures [Invited]

Xiaotao Guo, Xuwei Xue, Fulong Yan, Bitao Pan, Georgios Exarchakos, and Nicola Calabretta
J. Opt. Commun. Netw. 14(1) A69-A80 (2022)

Network-aware compute and memory allocation in optically composable data centers with deep reinforcement learning and graph neural networks

Zacharaya Shabka and Georgios Zervas
J. Opt. Commun. Netw. 15(2) 133-143 (2023)

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 (6)

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 (12)

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 (16)

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

	Definition
$N$	Set of nodes
$E$	Set of adjacent node pairs
$S^{c}$	Set of OCSs
$η_{s}^{c}$	Available port count in an OCS $s \in S^{c}$
$S^{p}$	Set of PSs
$η_{s}^{p}$	Available port count in a PS $s \in S^{p}$
$V$	Set of PS pairs
$κ_{g}$	Threshold of traffic flows in an optical path established of a PS pair $g \in V$
$J_{c}$	Set of resources corresponding to resource type $c$
$x$	Adjacency matrix representing node connection
$T$	Set of execution task types
$R_{t}$	Set of pairs of resource types requiring communication in task type $t \in T$
$A_{t}$	Proportion of type $t \in T$ tasks to total execution tasks
$H_{t, b, c}^{o}$	Maximum number of passed OCSs in resource type pair $b, c \in R_{t}$ for task type $t \in T$
$H_{t, b, c}^{p}$	Maximum number of passed PSs in resource type pair $b, c \in R_{t}$ for task type $t \in T$

Number	0	2	3	5	6	7
Code 1	1	0	0	1	0	1
Code 2	1	1	1	0	1	0
Generated code 1	1	0	1	0	1	1
Generated code 2	1	1	0	1	0	0

Parameters	Value
Clock speed of computational resources	3.4 GHz
FLOPS of the computational resources	76.8 GFLOPS
Memory processing delay	$0.125 µ s$
I/O delay	$0.101 µ s$
Page size	4 KB
Total computational resource pools	14
Total memory resource pools	2
Switching latency in EPSs	$1.11 µ s$
Ports of EPSs used to connect to other switches	20
Threshold of traffic flows in each EPS port	140
Switching latency in OCSs	$0.03 µ s$
Ports of OCSs used to connect to other switches	80
Threshold of traffic flows in each OPS port	140
Switching latency in OPSs	$0.55 µ s$
Ports of OPSs used to connect to other switches	64
Propagation delay (Switch-Switch)	$0.5 µ s$
Propagation delay (Resource-Switch)	$0.05 µ s$
The bandwidth of each optical fiber	10 Gbps

Parameter	Process 1	Process 2	Process 3
Clock count	0.035	0.054	2371.33
Packet rate to memory (/ms)	0.00033	0	1.87
Packet rate from memory (/ms)	0.00033	0.00033	3.71
Number of page faults	0	0	67,543.25
Number of pages per page fault	0	0	5.27
Memory resources	1	4 or 6	4 or 6
Computational resources	1	1	4 or 6

	Request 1	Request 2	Request 3
Acceptable time	400 ms	300 ms	200 ms
Maximum number of passed OCSs	6	3	3
Maximum number of passed OPSs	1	1	0
Proportion of each request generated to all requests (Case 1/2)	0.6/0.2	0.2/0.2	0.2/0.6

Resource Boards Connected to Each OCS	12	10
Computational resources connected to each OCS	576	480
Memory resources connected to each OCS	4608	3840
Total computational resources	8064	6720
Total memory resources	9216	7680

Parameter	Value
Iterations to generate physical topology with only OCSs	2000
Iterations to generate physical topology with OCSs and OPSs	2000
Topologies generated in each iteration	400
Proportions of select	0.1
Proportions of crossover	0.1
Proportions of mutation	0.8

	Case 1	Case 2
2D torus (16 OCSs + 1 PS)	0.5	0.5
3D torus (16 OCSs + 1 PS)	0.5	0.5
GT_env1 (16 OCSs + 1 PS)	1	1
GPT_env1 (16 OCSs + 1 PS)	0.8	0.4
Fat-tree (24 OCSs + 1 PS)	1	1
GT_env2 (24 OCSs + 1 PS)	1	1
GPT_env2 (24 OCSs + 1 PS)	0.9	0.7
2D torus (16 OCSs)	0.9	0.7
3D torus (16 OCSs)	0	0
GT_env3 (16 OCSs)	1	1
GPT_env3 (16 OCSs)	0.5	0.4
Fat-tree (24 OCSs)	1	1
GT_env4 (24 OCSs)	1	1
GPT_env4 (24 OCSs)	0.8	0.7

	Case 1		Case 2
	Average	Max	Average	Max
2D torus (16 OCSs + 1 PS)	4	5	4	5
3D torus (16 OCSs + 1 PS)	4.07	6	4.07	6
GT_env1 (16 OCSs + 1 PS)	3.57	4	3.50	4
GPT_env1 (16 OCSs + 1 PS)	3.71	5	3.79	5
Fat-tree (24 OCSs + 1 PS)	4	4	4	4
GT_env2 (24 OCSs + 1 PS)	3.68	4	3.46	4
GPT_env2 (24 OCSs + 1 PS)	3.75	5	3.89	5
2D torus (16 OCSs)	4	5	4	5
3D torus (16 OCSs)	4.07	6	4.07	6
GT_env3 (16 OCSs)	3.57	4	3.64	4
GPT_env3 (16 OCSs)	4.07	6	3.57	5
Fat-tree (24 OCSs)	4	4	4	4
GT_env4 (24 OCSs)	3.57	4	3.46	4
GPT_env4 (24 OCSs)	3.89	5	3.86	5

OCS	17	16	15	14	13	12	11	10	9	8	7	6	5	4	3	2	1	0
OPS	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17
CSTE (case 1)	0	1	1	1	1	1	1	1	1	1	1	1	1	0	0	0	0	0
CSTE (case 2)	0	1	1	1	1	1	1	1	1	1	1	1	0	0	0	0	0	0

	Case 1			Case 2
	1 Core	4 Cores	8 Cores	1 Core	4 Cores	8 Cores
Environment 1 (16 OCSs)
Cables	1156	308	114	1170	278	106
CSTE	1	1	1	1	1	1
Environment 2 (16 OCSs + 1 PS)
Cables	1244	318	150	1284	312	118
CSTE	1	1	0.5	1	1	0.5

Optical network topology design to execute many tasks simultaneously in a disaggregated data center

Author Affiliations

ORCID

Abstract

Cited By

Figures (6)

Tables (12)

Equations (16)

Journal of Optical Communications and Networking