idnits 2.17.1 draft-sun-nmrg-hybrid-switching-02.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** There is 1 instance of lines with control characters in the document. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == The document doesn't use any RFC 2119 keywords, yet seems to have RFC 2119 boilerplate text. -- The document date (Dec 19, 2020) is 1216 days in the past. Is this intentional? -- Found something which looks like a code comment -- if you have code sections in the document, please surround them with '' and '' lines. Checking references for intended status: Informational ---------------------------------------------------------------------------- No issues found here. Summary: 1 error (**), 0 flaws (~~), 2 warnings (==), 2 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Network Working Group W. Sun 3 Internet-Draft J. Shao 4 Intended status: Informational W. Hu 5 Expires: June 22, 2021 SJTU 6 Dec 19, 2020 8 Resource Allocation Model for Hybrid Switching Networks 9 draft-sun-nmrg-hybrid-switching-02.txt 11 Abstract 13 The fast increase in traffic volumn within and outside Datacenters is 14 placing an unprecendented challenge on the underline network, in both 15 the capacity it can provide, and the way it delivers traffic. When a 16 large portion of network traffic is contributed by large flows, 17 providing high capacity and slow to change optical circuit switching 18 along side fine-granular packet services may potentially improve 19 network utility and reduce both CAPEX and OpEX. This gives rise to 20 the concept of hybrid switching - a paradigm that seeks to make the 21 best of packet and circuit switching. 23 However, the full potential of hybrid switching networks (HSNs) can 24 only be realized when such a network is optimally designed and 25 operated, in the sense that "an appropriate amount of resource is 26 used to handle an appropriate amount of traffic in both switching 27 planes." The resource allocation problem in HSNs is in fact complex 28 ineractions between three components: resource allocation between the 29 two switching planes, traffic partitioning between the two switching 30 planes, and the overall cost or performance constraints. 32 In this memo, we explore the challenges of planning and operating 33 hybrid switching networks, with a particular focus on the resource 34 allocation problem, and provide a high-level model that may guide 35 resource allocation in future hybrid switching networks. 37 Status of This Memo 39 This Internet-Draft is submitted in full conformance with the 40 provisions of BCP 78 and BCP 79. 42 Internet-Drafts are working documents of the Internet Engineering 43 Task Force (IETF). Note that other groups may also distribute 44 working documents as Internet-Drafts. The list of current Internet- 45 Drafts is at https://datatracker.ietf.org/drafts/current/. 47 Internet-Drafts are draft documents valid for a maximum of six months 48 and may be updated, replaced, or obsoleted by other documents at any 49 time. It is inappropriate to use Internet-Drafts as reference 50 material or to cite them other than as "work in progress." 52 This Internet-Draft will expire on June 22, 2021. 54 Copyright Notice 56 Copyright (c) 2020 IETF Trust and the persons identified as the 57 document authors. All rights reserved. 59 This document is subject to BCP 78 and the IETF Trust's Legal 60 Provisions Relating to IETF Documents 61 (https://trustee.ietf.org/license-info) in effect on the date of 62 publication of this document. Please review these documents 63 carefully, as they describe your rights and restrictions with respect 64 to this document. 66 1. Introduction 68 In facing rapid increase of network traffic [Gantz12], as well as the 69 number of servers in cloud data centers [Cisco15], new architectures 70 and operation models of Data Center Networks (DCNs) gained wide 71 interests. One concept that attracted considerable and lasting 72 attention is the introduction of optical switching technologies into 73 DCNs, hoping that bypassing some of the traffic without performing 74 per-packet electronic processing will help reducing the Operational 75 Cost (OpEx), as well as the Capital Expenditure (CapEx) of DCNs. 76 This concept of combining electronic packet switching (EPS) and 77 optical switching (often optical circuit switching, OCS), is called 78 hybrid switching [Zukerman89]. In recent years, many hybrid 79 switching schemes have been proposed [Gauger06], and it is reasonable 80 to believe that when a DCN grows beyond a certain scale, the benefit 81 of introducing optical switching will emerge and become more evident 82 as the size of the DC continues to increase. 84 On the other hand, achieving the benefits of hybrid switching 85 requires careful design at the planning stage, and proper operation 86 during runtime. This poses challenges that goes far beyond the 87 topological or architectural aspects. For instance, at the planning 88 stage, one has to decide how much to invest in the two switching 89 planes, such that each could be fully utilized when the network 90 becomes operational. Under cases when dynamic resource allocation 91 between the two planes are possible, one has to decide how resource 92 is allocated between the two planes, and how traffic should be 93 directed to each of them, such that performance constraints can be 94 satisfied, and operational cost such as power consumption can be 95 minimized. 97 This memo aims to explore the challenges of planning and operating 98 hybrid switching networks, and provide a high-level model that may 99 guide the resource allocation in future hybrid switching networks. 100 We will use hybrid switching DCN as an example to show one possible 101 application of this model. 103 2. Conventions Used in This Document 105 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 106 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 107 document are to be interpreted as described in RFC2119. 109 3. Overview of Hybrid Switching Networks 111 Hybrid Switching Networks (HSNs) are networks that employ more than 112 one switching technology. The term started to attract attention when 113 Wavelength Dense Multiplexing (WDM) started to be deployed as a 114 underlying infrastructure of TCP/IP based packet networks [FENG17]. 115 It continued to receive considerable attention, as the research on 116 future-looking optical switching schemes boomed, around and after the 117 begining of the 21th century. 119 The research on hybrid switching gained momentum again with the rapid 120 growth of cloud data centers. With a clearer context and real-life 121 prototyping efforts, a wider concensus regarding the benefits and 122 feasibility of HSN have been reached. 124 The challenges of planning and operating hybrid DCNs are rooted in 125 the fundamental differences between EPS and OCS [Farrington10], 126 [WANG10]. In principle, EPS is good at delivering traffic that is 127 bursty and difficult to predict. By aggregating the traffic from a 128 large number of communcating peers, high network utilization can be 129 achieved at modest cost. OCS, on the other hand, is suited for well 130 planed, or highly predictable traffic patterns. One good example is 131 the delivery of bulk flows, which can last up to a few minutes when 132 carried by a wavelength channel at full capacity. 134 4. Terms used in this document 136 o Electronic Packet Switching (EPS) 137 EPS in this memo refers to the off-the-shelf switching technology. 138 It provides "best-effort" packet delivery service. Since EPS 139 performs fine-granular per-packet processing, it is generally 140 regarded to be best suitable for traffic that is bursty and 141 difficult to predict. Existing researches show that when lightly 142 loaded, the performance of EPS can be rather reliable and 143 predictable. However, when the network is heavily loaded, the 144 performance of EPS will deteriorate very quickly and result in 145 long queueing delay and high packet loss rate. 147 o Optical Circuit Switching (OCS) 148 OCS in this memo refers to connection oriented network services 149 based on optical switching technologies, such as MEMS or WSS based 150 switches, and the like. The connection oriented nature of OCS 151 requires the establishment of connections through signaling prior 152 to data transfer. The capacity of each connection, for instance, 153 a wavelength channel, often consumes a significant portion of the 154 overall network capacity. Request blocking is thus difficult to 155 eliminate in OCS, if not impossible. 157 o Hybrid Switching Networks (HSNs) 158 HSNs in this memo refers to networks that: i) employ both EPS and 159 OCS, and ii) accept data transfer request in both packet and 160 stream/flow form. Upon entering the network, requests in packets 161 form will be handled by the EPS plane, and requests in flow form 162 will be handled by OCS following the connection provisioning 163 procedures. This differs HSN from IP over WDM networks, where 164 both switching schemes exist, but services start and terminate 165 only on the IP layer, and standalone OCS service is not provided. 166 Note that the boundary between packet and flow requests may not 167 naturally exist. For instance, when flow level information is not 168 available from outside the network, it will be up to the network 169 to decide how traffic should be partitioned and then directed to 170 either EPS or OCS. 172 5. Performance Measures in Hybrid switching Networks 174 5.1. Performance Measures in Electronic Packet Switching 176 Without loss of generality, performance of packet switching networks 177 can be characterized by one or more of the following metrics: 179 o Packet loss rate - packet loss may happen when congestions occur. 180 Statistically, in a given network, packet loss rate can be seen as 181 a function of network load. Packet loss rarely happen when the 182 traffic load is low. But when the load increases to a certain 183 threshold in the network, or in part of it, packet loss rate may 184 increase quickly as load continues to increase. 185 Packet delay and jitter - like packet loss rate, packet delay is 186 mostly stable and jitter is small when the network is lightly 187 loaded. Delay and jitter will increase dramatically when network 188 load increases. 189 Flow completion time - flow completion time is a composite metric 190 that relies on both packet loss rate and packet delay. 192 5.2. Performance Measures in Optical Circuit Switching 194 Performance of Optical Circuit Switching (OCS) is typically measured 195 by request blocking rate, defined as the number of admitted requests 196 over the total number of arrivals. In theory, blocking in OCS can 197 not be eliminated. The planning of OCS is thus often a tradeoff 198 between performance and cost, as in the case of conventional 199 telephone networks, in which trunk capacity can be dimensioned with 200 the Erlang-B formula. 202 6. BLOC - the Blocking LOss Curves 204 6.1. General Idea 206 To understand the resource allocation in HSNs, it is important to 207 understand the interactions between the three components in the 208 systems: 210 o Traffic partitioning 211 Traffic partitioning means the separation of incoming traffic into 212 two parts so that each part can be handled by the one of the two 213 switching planes. In today's networks, there might be many 214 traffic separation/differentiation mechanisms for the purpose of 215 enforcing differentiated policy based on traffic type. Traffic 216 partitioning in the context of HSN, however, aims to realize the 217 optimal separation of flows into the two planes, such that the 218 utility of the network can be maximized. 219 One traffic partitioning method is a flow length based method. 220 With a predefined threshold, flows are classified into short flows 221 and large flows, each served with the packet switching plane and 222 the circuit switching plane, respectively. 223 Partitioning can be performed according to a priori knowledge, 224 e.g., according to the information provided by the applications 225 that generate the traffic flows. It also can be performed in 226 network during runtime. The details on traffic classification and 227 partitioning may be found in [Cisco15] and are outside the scope 228 of this memo. 230 o Resource allocation 231 The resource here can be physical resources such as switch ports, 232 wavelengths or fibers. It also can be abstract resource such as 233 the overall budget. 235 o Performance/Cost Constraints 236 The cost constraint applies when the making of the hybrid 237 switching system is subject to limited budget. For any given 238 traffic demand, the cost and performance of carrying the traffic 239 through either EPS or OCS can be very different. A good design 240 should, on the first hand, satisfy the performance constraint; on 241 the other, it should also leave space for future traffic demand 242 growth. The performance constraints specify the acceptable worst- 243 case performance of the system, for example, the maximum packet 244 loss rate, highest request-blocking rate, or longest packet delay 245 etc. Given traffic demand, the worst-case performance constraint 246 specifies the least amount of resource that should be allocated to 247 a switching plane. 249 As can easily be seen, the operation of HSNs involves close 250 interactions between the three components, and is a difficult 251 problem. The interactions can be summarized into the following 252 diagram. 254 +-----------------------+ 255 | Cost/Performance | 256 | Constraint | 257 +-----------------------+ 258 ^ 259 Constraint | Constraint 260 +---------------------+-------------------------+ 261 | | 262 v v 263 +-----------------+ +-----------------+ 264 | Traffic | Linkage | Resource | 265 | Partition |----------------------------| allocation | 266 | | | | 267 +-----------------+ +-----------------+ 269 Interactions beween the three components in HSN 271 6.2. Modeling the curves 273 In a typical IP network with a given traffic load, the packet loss 274 rate decreases when the network capacity increases and vice versa. 275 Similarly, in circuit switching networks, the request blocking 276 probability will decrease when the bandwidth increases and vice 277 versa. In a hybrid switching system, the overall resource capacity 278 is constant. The resource allocation between EPS and OCS plane will 279 directly affect the network performance of both switching planes. 280 The network performance is also affected directly by how the traffic 281 is partitioned between EPS and OCS planes. 283 ^ packet loss rate 284 | o capacity resource 1 285 | * capacity resource 2 ____o _____* ____x 286 | x capacity resource 3 / / / 287 | _____o _____* _____x 288 | / / / 289 | ____o ____* ____x 290 | / / / 291 | ___o ___* ___x 292 | / / / 293 | __o __* __x 294 | / / / 295 | _o _* _x 296 | / / / 297 | _o _* _x 298 | / / / 299 | _o _* _x 300 | / / / 301 | o * x 302 +--------------------------------------------------------> 303 the traffic load for EPS 305 Fig. 1(a) the performance curves for EPS 307 ^ request blocking rate 308 | o capacity resource 1 309 | * capacity resource 2 ____o _____* ____x 310 | x capacity resource 3 / / / 311 | _____o _____* _____x 312 | / / / 313 | ____o ____* ____x 314 | / / / 315 | ___o ___* ___x 316 | / / / 317 | __o __* __x 318 | / / / 319 | _o _* _x 320 | / / / 321 | _o _* _x 322 | / / / 323 | _o _* _x 324 | / / / 325 | o * x 326 +--------------------------------------------------------> 327 the traffic load for OCS 329 Fig. 1(b) the performance curves for OCS 331 To clearly classify the influence from the traffic load partition and 332 network capacity allocation, we take Fig. 1(a) and 1(b) to show the 333 performance curves for EPS and OCS with varying traffic loads and 334 network capacities. In Fig. 1(a), we choose the packet loss rate as 335 the performance of EPS. When the traffic load for EPS increases, the 336 packet loss rate becomes worse under the constrained network 337 capacity. The extension of network capacity will bring a promoted 338 packet loss rate for EPS which is classified in Fig. 1(a) with the 339 capacity resource 1 <= capacity resource 2 <= capacity resource 3. 340 The performance curves for OCS is shown in Fig. 1(b) after choosing 341 the request blocking rate as the y-axis, and the relationship among 342 these curves is still resource 1 <= capacity resource 2 <= capacity 343 resource 3. Combining Fig. 1(a) and 1(b), the more network 344 capacities we allocate to EPS or OCS, the better service they will 345 provide under a heavier traffic load transmission. 347 6.3. The BLOC System 349 The BLOC framework comprises two types of curves, i.e., loss curves 350 (LCs) and blocking curves (BCs), in the same two-dimensional 351 coordinate system. An LC or a BC in the BLOC framework is a curve 352 that contains a series of points with the same packet loss rate or 353 request blocking probability. Using the percentage of traffic 354 delivered by EPS as the x-axis and the percentage of bandwidth 355 allocated to EPS plane as the y-axis, all of the curves in the BLOC 356 framework are monotonic. Another important component of the BLOC 357 framework is the feasible region. In this paper, "feasible" means 358 that as long as the traffic partitioning and resource allocation fall 359 within this area, the resulting packet loss rate will be smaller than 360 Pmax (the maximum packet loss rate) and the request blocking 361 probability will be lower than Bmax (the maximum request blocking 362 rate). Thus, the feasible region contains all the feasible 363 combinations of resource allocation and traffic partitioning that 364 satisfy the network performance requirements. Different resource 365 allocation strategies in hybrid switching networks can be achieved by 366 choosing a point from the feasible region. 368 ^ 100% 369 |...% of resource allocated to the EPS plane (Y Axis)... 370 | _o _x 371 | (1) packet loss / / 372 | rate increase __o __x 373 | (2) packet loss / ^ / 374 | rate decrease _o . _x 375 | / . / 376 | __o .(2) __x 377 | / . / 378 | __o . __x 379 | / . / 380 | ___o . ___x 381 | / (1) ./ 382 | ____o .............>____x 383 | / / 384 | _____o ____x 385 | / / 386 |_o____________________x 387 +--------------------------------------------------------> 388 (0,0) % of traffic offered to the EPS plane (X Axis) 100% 390 Fig. 2(a) the packet loss curves for EPS plane 392 ^ 100% 393 | (100%,100%) 394 |...% of resource allocated to the EPS plane (Y Axis)... + 395 | o______________x______/ 396 | / / 397 | _____o _____x 398 | / / 399 | ____o ____x 400 | / ^ / 401 | ___o . ___x 402 | / (2). / 403 | __o . __x 404 | / (1) ./ 405 | __o ........> __x (1) request blocking 406 | / / rate decrease 407 | _o _x (2) request blocking 408 | / / rate increase 409 | _o _x 410 | / / 411 |o x 412 +--------------------------------------------------------> 413 % of traffic offered to the EPS plane (X Axis) 100% 415 Fig. 2(b) the request blocking curves for OCS plane 417 Fig. 2(a) and 2(b) show an example of LCs and BCs when the overall 418 hybrid system capacity and the traffic volume are fixed. In Fig. 419 2(a), when the percentage of traffic to be transmitted by EPS 420 increases, the bandwidth allocated to EPS plane must also be 421 increased so the same packet loss rate can be achieved. Hence, each 422 LC is monotonically increasing. In addition, the LCs with smaller 423 loss rate values require a larger percentage of bandwidth for the 424 same amount of traffic. Therefore, the LCs moves to the top left 425 when the packet loss rate becomes smaller, as shown in Fig. 2(a). 426 All of the LCs pass through the origin (0, 0), so if no bandwidth is 427 allocated to PS plane, it cannot transmit any traffic. Similarly, 428 the BCs move downward to the right when the request blocking 429 probability becomes lower, and all of the BCs converge to the point 430 (100,100), where all of the bandwidth and traffic is assigned to PS 431 plane, as shown in Fig. 2(b). 433 ^ 100% 434 | 435 |----- % of resource allocated to the EPS plane (Y Axis) 436 | +----------------------------------+ 437 | | Request blocking curve with Bmax | / 438 | +------------------+---------------+ / 439 | | / 440 | +-----------------------------+ | / / 441 | | Packet loss curve with Pmax | | / / 442 | +------+----------------------+ | / / 443 | | v / / 444 | | /-------------------O---------X-/ 445 | | / -- -- -- -- -- -- -- -- -- / 446 | | / -- -- -- -- -- -- -- -- -- / 447 | v / -- -- -- -- -- -- -- -- -- / 448 | /-O-.-----------------^-----------/ 449 | / / | 450 | / / | 451 | / / +---------------------+ 452 | / / | feasible region | 453 | / +---------------------+ 454 | / 455 +--------------------------------------------------------> 456 % of traffic offered to the EPS plane (X Axis) 100% 458 Fig. 3 an examplery BLOC 460 We now consider a hybrid switching system with the maximal allowed 461 packet loss rate Pmax and the maximal allowed request blocking 462 probability Bmax [FENG16]. Fig. 3 shows a BLOC where the LCs and BCs 463 are placed in the same two-dimensional coordinate system. The 464 hatched area above the LC of Pmax and below the BC of Bmax contains 465 all of the feasible combinations of traffic partitioning and resource 466 allocation. Choosing a point from the feasible region (i.e., a 467 combination of resource allocation and traffic partitioning) is 468 subject to various optimization objectives. For instance, from an 469 energy consumption perspective, we need to choose the point with the 470 minimal percentage of EPS resources from the feasible region (i.e., 471 the lowest point in the feasible region), so that the overall energy 472 consumption would be minimized. In Section 5, we show that other 473 metrics can also be optimized with the BLOC, such as the packet delay 474 in EPS plane as a function of resource allocation and traffic 475 partitioning. 477 6.4. An example 479 A hybrid switching Datacenter network is shown in Fig. 4 [FENG17]. 480 Among all s+p uplink interfaces on each ToR switch, s of them connect 481 the switch to the EPS plane and the rest, p, connect the ToR switch 482 to the OCS network. As the cost of supporting an OCS connection can 483 be very different from that of an EPS port, different combinations of 484 s and p will result in significant difference in building cost. 485 Different combinations of s and p will also lead to different 486 performance and running cost, such as power consumption. 488 +-------------------------+ +-----------------------+ 489 | | | | 490 | EPS Network | | OCS Network | 491 | +-+ +-+ +-+ +-+ +-+ +-+ | |+-+ +-+ +-+ +-+ +-+ +-+| 492 +-+-+-+-+-+-+-+-+-+-+-+-+-+ ++-+-+-+-+-+-+-+-+-+-+-++ 493 ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ 494 | | | | \ \ / | / / \ \ 495 | | | | \ \ / / / / \ \ 496 | | | | \ \ / / / / | | 497 | | | \ \ \-------*-----*-*---*---\ | | 498 | | | \ \---------*-----*-*---*\ \ \ \ 499 .+---+. \ *---------------/ / / / \ \ \ \ 500 s -->( | | ) \ / \ /-----------------/ / / \ \ | \ 501 `+---+' \ / *-------\ / / \ \ | | 502 | | *---*-----\ \ / / \ | | | 503 | | .X--./ \ \ / / | | | | 504 p -----+---+-->(/ /) \ \ / / | | | | 505 v v v`---' v v v v v v v v 506 +-+-+-+-+-+-+-+-+-+ +-+-+-+-+-+-+-+-+-+ +-+-+-+-+-+-+-+-+-+ 507 | +-+ +-+ +-+ +-+ | | +-+ +-+ +-+ +-+ | | +-+ +-+ +-+ +-+ | 508 | ToR Switch | | ToR Switch | | ToR Switch | 509 +-----------------+ +-----------------+ +-----------------+ 510 +-----------------+ +-----------------+ +-----------------+ 511 +-----------------+ +-----------------+ +-----------------+ 512 | ... | | ... | | ... | 513 +-----------------+ +-----------------+ +-----------------+ 514 +-----------------+ +-----------------+ +-----------------+ 515 +-----------------+ +-----------------+ +-----------------+ 517 Fig. 4 switch ports allocation in hybrid DCN 519 The costs of network interconnecting devices in the EPS and OCS 520 networks are determined by allocation of uplink interfaces. Thus, 521 for each ToR, the cost constraint can be presented as Cp(s) + Cc(p)<= 522 C, in which Cp(s) stands for the cost of EPS with s uplinks, and 523 Cc(p) stands for the OCS cost with p uplinks. 525 The total volume of flows to be transmitted on a ToR switch is V. 526 The traffic is carried by either EPS or OCS:Vp + Vc = V. 528 The performance requirements specify the acceptable worst-case 529 performance of the system, such as the longest flow completion time 530 and the highest request blocking probability. A proper resource 531 allocation and traffic partitioning should satisfy the performance 532 requirements in both EPS and OCS networks: Tp(s,Vp) <= Tmax, Bc(p,Vc) 533 <= Bmax, where Tmax and Bmax are the flow completion time and request 534 blocking probability requirements in EPS and OCS, respectively. 536 ^ 100% 537 | % of resource to EPS 538 | +------------+ +-----------+ * 539 | |Tmax=0.02ms | | Bmax=0.01 | * 540 | +------------+ +-----------+ **---- 541 | | | */ //// 542 | | | */ //// 543 | | | */ ///// 544 | v ***********************/ ////// 545 | ***O******** v / / +++++ 546 | ** /-----------------O----------/ / + 547 | ** / ////////////////////////// / / + 548 | * / ////////////////////////// / / + 549 | * /--------O-------------------/ / + 550 | * / ////////^////////////////////// + 551 | ** / /////////|///////////////////// + 552 | * / // +++++++|+++++++++++++O++++++++ 553 | * /+++++ | ^ 554 | * /+ | | 555 | * ++ +-----------+ +-----------+ 556 | * + |Bmax=0.001 | | Tmax=1ms | 557 | ** ++ +-----------+ +-----------+ 558 | ***+++ 559 | *** 100% 560 +--------------------------------------------------------> 561 % of traffic offered to the EPS 563 Fig. 5 BLOC for hybrid DCN 565 Fig. 5 shows the BLOC with different network performance 566 requirements. When Tmax equals 1 ms and Bmax equals 0.01, there is a 567 feasible region between the curves with Tmax and Bmax. When the 568 performance requirements are higher (i.e., smaller Tmax and Bmax), 569 the feasibleregion will be smaller or may even disappear. For 570 example, when Tmax and Bmax decrease respectively to 0.02 ms and 571 0.001, the feasible region cannot be found. That means it is not 572 possible to find a resource allocation that can satisfy Tmax and Bmax 573 simultaneously. As the hybrid switching system is an interaction 574 between the three components, when the network performance 575 requirements cannot be satisfied, the system should have a greater 576 budget or carry less traffic to obtain a feasible resource 577 allocation. 579 7. Security Considerations 581 This document does not impose any new challenges to the current 582 Internet. 584 8. IANA Considerations 586 This document makes no requests for IANA action. 588 9. Acknowledgements 590 We are grateful to the valuable discussions and inputs from the 591 community. We thank the support from NSFC. 593 10. Informative References 595 [Cisco15] Cisco, Cisco., "Cisco global cloud index: Forecast and 596 methodology, 2015-2020. white paper", http://www.cisco.com 597 /en/US/solutions/collateral/ns341/ns525/ns537/ns705/ 598 ns1175/Cloud_Index_White_Paper.html#wp9000816 1-29, 2015. 600 [Farrington10] 601 Farrington, Nathan., Porter, George., Radhakrishnan, 602 Sivasankar., Bazzaz,, Hamid., Subramanya, Vikram., 603 Fainman, Yeshaiahu., Papen, George., and Amin. Vahdat, 604 "Helios: a hybrid electrical/optical switch architecture 605 for modular data centers", SIGCOMM'10 339-350, 606 DOI 10.1145/1851182.1851223, August 2010. 608 [FENG16] Feng, Z., Sun, W., and W. Hu, "BLOC: A Generic Resource 609 Allocation Framework for Hybrid Packet/Circuit-Switched 610 Networks", J. Opt. Commun. Netw. 8, 689-700, 611 DOI 10.1364/JOCN.8.000689, August 2016. 613 [FENG17] Feng, Z., Sun, W., Zhu, J., Shao, J., and W. Hu, "Resource 614 Allocation in Electrical/Optical Hybrid Switching Data 615 Center Networks", J. Opt. Commun. Netw. 9, 648-657, 616 DOI 10.1364/JOCN.8.000689, August 2017. 618 [Gantz12] Gantz, John. and David. Reinsel, "The digital universe in 619 2020: Big data, bigger digital shadows, and biggest growth 620 in the far east", International Data Corporation 1414_v2, 621 December 2012. 623 [Gauger06] 624 Gauger, C., Kuhn, P., Breusegem, E., Pickavet, M., and P. 625 Demeester, "Hybrid optical network architectures: Bringing 626 packets and circuits together", IEEE Commun. Mag. 44(8), 627 36-42, DOI 10.1109/MCOM.2006.1678107, August 2006. 629 [WANG10] Wang, Guohui., Andersen, David., Kaminsky, Michael., 630 Papagiannaki, Konstantina., Eugene Ng, T., Kozuch, 631 Michael., and Michael. Ryan, "c-Through: part-time optics 632 in data centers", SIGCOMM'10 327-338, 633 DOI 10.1145/1851182.1851223, August 2010. 635 [Zukerman89] 636 Zukerman, M., "Bandwidth allocation for bursty isochronous 637 traffic in a hybrid switching system", IEEE Transactions 638 on Communications 37(12), 1367-1371, DOI 10.1109/26.44208, 639 December 1989. 641 Authors' Addresses 643 Weiqiang Sun 644 Shanghai Jiao Tong University 645 800 Dongchuan Road 646 Shanghai 200240 647 China 649 Phone: +86 21 3420 5359 650 EMail: sun.weiqiang@gmail.com 652 Junyi Shao 653 Shanghai Jiao Tong University 654 800 Dongchuan Road 655 Shanghai 200240 656 China 658 EMail: shaojunyi@sjtu.edu.cn 660 Weisheng Hu 661 Shanghai Jiao Tong University 662 800 Dongchuan Road 663 Shanghai 200240 664 China 666 EMail: wshu@sjtu.edu.cn