idnits 2.17.1 draft-ietf-roll-trickle-05.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 : ---------------------------------------------------------------------------- No issues found here. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (November 10, 2010) is 4917 days in the past. Is this intentional? Checking references for intended status: Proposed Standard ---------------------------------------------------------------------------- (See RFCs 3967 and 4897 for information about using normative references to lower-maturity documents in RFCs) == Outdated reference: A later version (-01) exists of draft-hui-6man-trickle-mcast-00 == Outdated reference: A later version (-19) exists of draft-ietf-roll-rpl-15 Summary: 0 errors (**), 0 flaws (~~), 3 warnings (==), 1 comment (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Networking Working Group P. Levis 3 Internet-Draft Stanford University 4 Intended status: Standards Track T. Clausen 5 Expires: May 14, 2011 LIX, Ecole Polytechnique 6 J. Hui 7 Arch Rock Corporation 8 O. Gnawali 9 Stanford University 10 J. Ko 11 Johns Hopkins University 12 November 10, 2010 14 The Trickle Algorithm 15 draft-ietf-roll-trickle-05 17 Abstract 19 The Trickle algorithm allows nodes in a lossy shared medium (e.g., 20 low power and lossy networks) to exchange information in a highly 21 robust, energy efficient, simple, and scalable manner. Dynamically 22 adjusting transmission windows allows Trickle to spread new 23 information on the scale of link-layer transmission times while 24 sending only a few messages per hour when information does not 25 change. A simple suppression mechanism and transmission point 26 selection allows Trickle's communication rate to scale 27 logarithmically with density. This document describes the Trickle 28 algorithm and considerations in its use. 30 Status of this Memo 32 This Internet-Draft is submitted in full conformance with the 33 provisions of BCP 78 and BCP 79. 35 Internet-Drafts are working documents of the Internet Engineering 36 Task Force (IETF). Note that other groups may also distribute 37 working documents as Internet-Drafts. The list of current Internet- 38 Drafts is at http://datatracker.ietf.org/drafts/current/. 40 Internet-Drafts are draft documents valid for a maximum of six months 41 and may be updated, replaced, or obsoleted by other documents at any 42 time. It is inappropriate to use Internet-Drafts as reference 43 material or to cite them other than as "work in progress." 45 This Internet-Draft will expire on May 14, 2011. 47 Copyright Notice 48 Copyright (c) 2010 IETF Trust and the persons identified as the 49 document authors. All rights reserved. 51 This document is subject to BCP 78 and the IETF Trust's Legal 52 Provisions Relating to IETF Documents 53 (http://trustee.ietf.org/license-info) in effect on the date of 54 publication of this document. Please review these documents 55 carefully, as they describe your rights and restrictions with respect 56 to this document. Code Components extracted from this document must 57 include Simplified BSD License text as described in Section 4.e of 58 the Trust Legal Provisions and are provided without warranty as 59 described in the Simplified BSD License. 61 Table of Contents 63 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3 64 2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 3 65 3. Trickle Algorithm Overview . . . . . . . . . . . . . . . . . . 4 66 4. Trickle Algorithm . . . . . . . . . . . . . . . . . . . . . . 5 67 4.1. Parameters and Variables . . . . . . . . . . . . . . . . . 5 68 4.2. Algorithm Description . . . . . . . . . . . . . . . . . . 6 69 5. Using Trickle . . . . . . . . . . . . . . . . . . . . . . . . 6 70 6. Operational Considerations . . . . . . . . . . . . . . . . . . 7 71 6.1. Mismatched redundancy constants . . . . . . . . . . . . . 7 72 6.2. Mismatched Imin . . . . . . . . . . . . . . . . . . . . . 7 73 6.3. Mismatched Imax . . . . . . . . . . . . . . . . . . . . . 7 74 6.4. Mismatched definitions . . . . . . . . . . . . . . . . . . 8 75 6.5. Specifying the constant k . . . . . . . . . . . . . . . . 8 76 6.6. Relationship between k and Imin . . . . . . . . . . . . . 8 77 6.7. Tweaks and improvements to Trickle . . . . . . . . . . . . 9 78 6.8. Uses of Trickle . . . . . . . . . . . . . . . . . . . . . 9 79 7. Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 10 80 8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 10 81 9. Security Considerations . . . . . . . . . . . . . . . . . . . 10 82 10. References . . . . . . . . . . . . . . . . . . . . . . . . . . 11 83 10.1. Normative References . . . . . . . . . . . . . . . . . . . 11 84 10.2. Informative References . . . . . . . . . . . . . . . . . . 11 85 Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . . 12 87 1. Introduction 89 The Trickle algorithm establishes a density-aware local communication 90 primitive with an underlying consistency model that guides when a 91 node transmits. When a node's data does not agree with its 92 neighbors, that node communicates quickly to resolve the 93 inconsistency (e.g., in milliseconds). When nodes agree, they slow 94 their communication rate exponentially, such that nodes send packets 95 very infrequently (e.g., a few packets per hour). Instead of 96 flooding a network with packets, the algorithm controls the send rate 97 so each node hears a small trickle of packets, just enough to stay 98 consistent. Furthermore, by relying only on local communication 99 (e.g., broadcast or local multicast), Trickle handles network re- 100 population, is robust to network transience, loss, and disconnection, 101 is simple to implement, and requires very little state. Current 102 implementations use 4-11 bytes of RAM and are 50-200 lines of C 103 code[Levis08]. 105 While Trickle was originally designed for reprogramming protocols 106 (where the data is the code of the program being updated), experience 107 has shown it to be a powerful mechanism that can be applied to wide 108 range of protocol design problems, including control traffic timing, 109 multicast propagation, and route discovery. This flexibility stems 110 from being able to define, on a case-by-case basis, what constitutes 111 "agreement" or an "inconsistency;" Section Section 6.8 presents a few 112 examples of how the algorithm can be used. 114 This document describes the Trickle algorithm and provides guidelines 115 for its use. It also states requirements for protocol specifications 116 that use Trickle. This document does not provide results on 117 Trickle's performance or behavior, nor does it explain the 118 algorithm's design in detail: interested readers should refer to 119 [Levis04] and [Levis08]. 121 2. Terminology 123 The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 124 "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and 125 "OPTIONAL" in this document are to be interpreted as described in RFC 126 2119 [RFC2119]. 128 Additionally, this document introduces the following terminology: 130 Trickle communication rate: the sum of the number of messages sent 131 or received by the Trickle algorithm in an interval. 133 Trickle transmission rate: the sum of the number of messages sent by 134 the Trickle algorithm in an interval. 136 3. Trickle Algorithm Overview 138 Trickle's basic primitive is simple: every so often, a node transmits 139 data unless it hears a few other transmissions whose data suggest its 140 own transmission is redundant. Examples of such data include routing 141 state, software update versions, and the last heard multicast packet. 142 This primitive allows Trickle to scale to thousand-fold variations in 143 network density, quickly propagate updates, distribute transmission 144 load evenly, be robust to transient disconnections, handle network 145 repopulations, and impose a very low maintenance overhead: one 146 example use, routing beacons in the CTP protocol [Gnawali09], 147 requires sending on the order of a few packets per hour yet can 148 respond in milliseconds. 150 Trickle sends all messages to a local communication address. The 151 exact address used can depend on both the underlying IP protocol as 152 well as how the higher layer protocol uses Trickle. In IPv6, for 153 example, it can be the link-local multicast address or another local 154 multicast address, while in IPv4 it can be the broadcast address 155 (255.255.255.255). 157 There are two possible results to a Trickle message: either every 158 node that hears the message finds its data is consistent with their 159 own state, or a recipient detects an inconsistency. Detection can be 160 the result of either an out-of-date node hearing something new, or an 161 updated node hearing something old. As long as every node 162 communicates somehow - either receives or transmits - some node will 163 detect the need for an update. 165 For example, consider a simple case where "up to date" is defined by 166 version numbers (e.g., network configuration). If node A transmits 167 that it has version V, but B has version V+1, then B knows that A 168 needs an update. Similarly, if B transmits that it has version V+1, 169 A knows that it needs an update. If B broadcasts or multicasts 170 updates, then all of its neighbors can receive them without having to 171 advertise their need. Some of these recipients might not even have 172 heard A's transmission. 174 In this example, it does not matter who first transmits, A or B; 175 either case will detect the inconsistency. All that matters is that 176 some nodes communicate with one another at some nonzero rate. As 177 long as the network is connected and there is some minimum 178 communication rate for each node, the network will reach eventual 179 consistency. 181 The fact that Trickle communication can be either transmission or 182 reception enables the Trickle algorithm to operate in sparse as well 183 as dense networks. A single, disconnected node must transmit at the 184 Trickle communication rate. In a lossless, single-hop network of 185 size n, the Trickle communication rate at each node equals the sum of 186 the Trickle transmission rates across all nodes. The Trickle 187 algorithm balances the load in such a scenario, as each node's 188 Trickle transmission rate is 1/nth of the Trickle communication rate. 189 Sparser networks require more transmissions per node, but the 190 utilization of a given broadcast domain (e.g., radio channel over 191 space, shared medium) will not increase. This is an important 192 property in wireless networks and other shared media, where the 193 channel is a valuable shared resource. Additionally, reducing 194 transmissions in dense networks conserves system energy. 196 4. Trickle Algorithm 198 This section describes the Trickle algorithm. 200 4.1. Parameters and Variables 202 A Trickle timer has three configuration parameters: the minimum 203 interval size Imin, the maximum interval size Imax, and a redundancy 204 constant k: 206 o The minimum interval size, Imin, is defined in units of time 207 (e.g., milliseconds, seconds). For example, a protocol might 208 define the minimum interval as 100 milliseconds. 210 o The maximum interval size, Imax, is described as a number of 211 doublings of the minimum interval size (the base-2 log(max/min)). 212 For example, a protocol might define Imax as 16. If the minimum 213 interval is 100ms, then the amount of time specified by Imax is 214 100ms * 65536, 6,553.6 seconds, or approximately 109 minutes. 216 o The redundancy constant is a natural number (an integer greater 217 than zero). 219 In addition to these three parameters, Trickle maintains three 220 variables: 222 o I, the current interval size 224 o t, a time within the current interval, and 226 o c, a counter. 228 4.2. Algorithm Description 230 The Trickle algorithm has five rules: 232 1. When an interval begins, Trickle resets c to 0 and sets t to a 233 random point in the interval, taken from the range [I/2, I), that 234 is, values greater than or equal to I/2 and less than I. The 235 interval ends at I. 237 2. Whenever Trickle hears a transmission that is "consistent," it 238 increments the counter c. 240 3. At time t, Trickle transmits if and only if the counter c is less 241 than the redundancy constant k. 243 4. When the interval I expires, Trickle doubles the interval length. 244 If this new interval length would be longer than the time 245 specified by Imax, Trickle sets the interval length I to be the 246 time specified by Imax. 248 5. If Trickle hears a transmission that is "inconsistent," the 249 Trickle timer resets. If I is greater than Imin when a Trickle 250 timer resets, Trickle sets I to Imin, resets c to 0, and sets t 251 to random point in the interval, taken from the range [I/2, I), 252 that is, values greater than or equal to I/2 and less than I. If 253 I is equal to Imin, resetting a Trickle timer does nothing. 254 Trickle can also reset the timer in response to external 255 "events." 257 The terms consistent, inconsistent and event are in quotes because 258 their meaning depends on how a protocol uses Trickle. 260 5. Using Trickle 262 A protocol specification that uses Trickle MUST specify: 264 o Default values for Imin, Imax, and k. Because link layers can 265 vary widely in their properties, the default value of Imin SHOULD 266 be specified in terms of the worst-case latency of a link layer 267 transmission. For example, a specification should say "the 268 default value of Imin is 4 times the worst case link layer 269 latency" and should not say "the default value of Imin is 500 270 milliseconds." Worst case latency is approximately time until the 271 first link-layer transmission of the frame assuming an idle 272 channel (does not include backoff, virtual carrier sense, etc.). 274 o What constitutes a "consistent" transmission. 276 o What constitutes an "inconsistent" transmission. 278 o What "events," if any, besides inconsistent transmissions that 279 reset the Trickle timer. 281 6. Operational Considerations 283 It is RECOMMENDED that a protocol which uses Trickle include 284 mechanisms to inform nodes of configuration parameters at runtime. 285 However, it is not always possible to do so. In the cases where 286 different nodes have different configuration parameters, Trickle may 287 have unintended behaviors. This section outlines some of those 288 behaviors and operational considerations as educational exercises. 290 6.1. Mismatched redundancy constants 292 If nodes do not agree on the redundancy constant k, then nodes with 293 higher values of k will transmit more often than nodes with lower 294 values of k. In some cases, this increased load can be independent 295 of the density. For example, consider a network where all nodes but 296 one have k=1, and this one node has k=2. The different node can end 297 up transmitting on every interval: it is maintaining a Trickle 298 communication rate of 2 with only itself. Hence, the danger of 299 mismatched k values is uneven transmission load that can deplete the 300 energy of some nodes in a low power network. 302 6.2. Mismatched Imin 304 If nodes do not agree on Imin, then some nodes, on hearing 305 inconsistent messages, will transmit sooner than others. These 306 faster nodes will have their intervals grow to similar size as the 307 slower nodes within a single slow interval time, but in that period 308 may suppress the slower nodes. However, such suppression will end 309 after the first slow interval, when the nodes generally agree on the 310 interval size. Hence, mismatched Imin values are usually not a 311 significant concern. Note that mismatched Imin values and matching 312 Imax doubling constants will lead to mismatched maximum interval 313 lengths. 315 6.3. Mismatched Imax 317 If nodes do not agree on Imax, then this can cause long-term problems 318 with transmission load. Nodes with small Imax values will transmit 319 faster, suppressing those with larger Imax values. The nodes with 320 larger Imax values, always suppressed, will never transmit. In the 321 base case, when the network is consistent, this can cause long-term 322 inequities in energy cost. 324 6.4. Mismatched definitions 326 If nodes do not agree on what constitutes a consistent or 327 inconsistent transmission, then Trickle may fail to operate properly. 328 For example, if a receiver thinks a transmission is consistent, but 329 the transmitter (if in the receivers situation) would have thought it 330 inconsistent, then the receiver will not respond properly and inform 331 the transmitter. This can lead the network to not reach a consistent 332 state. For this reason, unlike the configuration constants k, Imin, 333 and Imax, consistency definitions MUST be clearly stated in the 334 protocol and SHOULD NOT be configured at runtime. 336 6.5. Specifying the constant k 338 There are some edge cases where a protocol may wish to use Trickle 339 with its suppression disabled (k is set to infinity). In general, 340 this approach is highly dangerous and it is NOT RECOMMENDED. 341 Disabling suppression means that every node will always send on every 342 interval, and can lead to congestion in dense networks. This 343 approach is especially dangerous if many nodes reset their intervals 344 at the same time. In general, it is much more desirable to set k to 345 a high value (e.g., 5 or 10) than infinity. Typical values for k are 346 1-5: these achieve a good balance between redundancy and low 347 cost[Levis08]. 349 Nevertheless, there are situations where a protocol may wish to turn 350 off Trickle suppression. Because k is a natural number 351 (Section 4.1), k=0 has no useful meaning. If a protocol allows k to 352 be dynamically configured, a value of 0 remains unused. For ease of 353 debugging and packet inspection, having the parameter describe k-1 354 rather than k can be confusing. Instead, it is RECOMMENDED that 355 protocols which require turning off suppression reserve k=0 to mean 356 k=infinity. 358 6.6. Relationship between k and Imin 360 Finally, a protocol SHOULD set k and Imin such that Imin is at least 361 two to three times as long as it takes to transmit k packets. 362 Otherwise, if more than k nodes reset their intervals to Imin, the 363 resulting communication will lead to congestion and significant 364 packet loss. Experimental results have shown that packet losses from 365 congestion reduce Trickle's efficiency [Levis04]. 367 6.7. Tweaks and improvements to Trickle 369 Trickle is based on a small number of simple, tightly integrated 370 mechanisms that are highly robust to challenging network 371 environments. In our experiences using Trickle, attempts to tweak 372 its behavior are typically not worth the cost. As written, the 373 algorithm is already highly efficient: further reductions in 374 transmissions or response time come at the cost of failures in edge 375 cases. Based on our experiences, we urge protocol designers to 376 suppress the instinct to tweak or improve Trickle without a great 377 deal of experimental evidence that the change does not violate its 378 assumptions and break the algorithm in edge cases. 380 This warning in mind, Trickle is far from perfect. For example, 381 Trickle suppression typically leads sparser nodes to transmit more 382 than denser ones; it is far from the optimal computation of a minimum 383 cover. However, in dynamic network environments such as wireless and 384 low-power, lossy networks, the coordination needed to compute the 385 optimal set of transmissions is typically much greater than the 386 benefits it provides. One of the benefits of Trickle is that it is 387 so simple to implement and requires so little state yet operates so 388 efficiently. Efforts to improve it should be weighed against the 389 cost of increased complexity. 391 6.8. Uses of Trickle 393 The Trickle algorithm has been used in a variety of protocols, both 394 in operational as well as academic settings. Giving a brief overview 395 of some of these uses provides useful examples of how and when it can 396 be used. These examples should not be considered exhaustive. 398 Reliable flooding/dissemination: A protocol uses Trickle to 399 periodically advertise the most recent data it has received, 400 typically through a version number. An inconsistency is when a node 401 hears a newer version number or receives new data. A consistency is 402 when a node hears an older or equal version number. When hearing an 403 older version number, rather than reset its own Trickle timer, it 404 sends an update. Nodes with old version numbers that receive the 405 update will then reset their own timers, leading to fast propagation 406 of the new data. Examples of this use include 407 multicast[I-D.hui-6man-trickle-mcast], network 408 configuration[Lin08][Dang09], and installing new application 409 programs[Hui04][Levis04]. 411 Routing control traffic: A protocol uses Trickle to control when it 412 sends beacons which contain routing state. An inconsistency is when 413 the routing topology changes in a way that could lead to loops or 414 significant stretch: examples include when the routing layer detects 415 a routing loop or when a node's routing cost changes significantly. 416 Consistency is when the routing topology is operating well and is 417 delivering packets successfully. Using the Trickle algorithm in this 418 way allows a routing protocol to react very quickly to problems (Imin 419 is small) but send very few beacons when the topology is stable. 420 Examples of this use include RPL[I-D.ietf-roll-rpl], CTP[Gnawali09], 421 and some current commericial IPv6 routing layers[Hui08]. 423 7. Acknowledgements 425 The authors would like to acknowledge the guidance and input provided 426 by the ROLL chairs, David Culler and JP Vasseur. 428 The authors would also like to acknowledge the helpful comments of 429 Yoav Ben-Yehezkel, Alexandru Petrescu, and Urlich Herberg, which 430 greatly improved the document. 432 8. IANA Considerations 434 This document has no IANA considerations. 436 9. Security Considerations 438 As it is an algorithm, Trickle itself does not have any specific 439 security considerations. However, two security concerns can arise 440 when Trickle is used in a protocol. The first is that an adversary 441 can force nodes to send many more packets than needed by forcing 442 Trickle timer resets. In low power networks this increase in traffic 443 can harm system lifetime. The second concern is that an adversary 444 can prevent nodes from reaching consistency. 446 Protocols can prevent adversarial Trickle resets by carefully 447 selecting what can cause a reset and protecting these events and 448 messages with proper security mechanisms. For example, if a node can 449 reset nearby Trickle timers by sending a certain packet, this packet 450 should be authenticated such that an adversary cannot forge one. 452 An adversary can possibly prevent nodes from reaching consistency by 453 suppressing transmissions with "consistent" messages. For example, 454 imagine node A detects an inconsistency and resets its Trickle timer. 455 If an adversary can prevent A from sending messages that inform 456 nearby nodes of the inconsistency in order to repair it, then A may 457 remain inconsistent indefinitely. Depending on the security model of 458 the network, authenticated messages, or a transitive notion of 459 consistency can prevent this problem. E.g., if messages that are 460 consistent with A and so suppress its transmissions are by definition 461 inconsistent with what A heard, then an adversary cannot 462 simultaneously prevent A from notifying neighbors and not notify the 463 neighbors itself (recall Trickle operates on shared, broadcast 464 media). Note that this means Trickle should filter unicast messages. 466 10. References 468 10.1. Normative References 470 [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate 471 Requirement Levels", BCP 14, RFC 2119, March 1997. 473 10.2. Informative References 475 [Dang09] Dang, T., Bulusu, N., Feng, W., and S. Park, "DHV: A Code 476 Consistency Maintenance Protocol for Multi-hop Wireless 477 Networks", Wireless Sensor Networks: 6th European 478 Conference Proceedings EWSN 2009 Cork, February 2009, 479 . 481 [Gnawali09] 482 Gnawali, O., Fonseca, R., Jamieson, K., Moss, D., and P. 483 Levis, "Collection Tree Protocol", Proceedings of the 7th 484 ACM Conference on Embedded Networked Systems SenSys 2009, 485 November 2009, 486 . 488 [Hui04] Hui, J. and D. Culler, "The dynamic behavior of a data 489 dissemination protocol for network programming at scale", 490 Proceedings of the 2nd ACM Conference on Embedded 491 Networked Systems SenSys 2004, November 2004, 492 . 494 [Hui08] Hui, J. and D. Culler, "IP is dead, long live IP for 495 wireless sensor networks", Proceedings of the 6th ACM 496 Conference on Embedded Networked Systems SenSys 2008, 497 November 2008, 498 . 500 [I-D.hui-6man-trickle-mcast] 501 Hui, J. and R. Kelsey, "Multicast Forwarding Using 502 Trickle", draft-hui-6man-trickle-mcast-00 (work in 503 progress), July 2010. 505 [I-D.ietf-roll-rpl] 506 Winter, T., Thubert, P., Brandt, A., Clausen, T., Hui, J., 507 Kelsey, R., Levis, P., Pister, K., Struik, R., and J. 508 Vasseur, "RPL: IPv6 Routing Protocol for Low power and 509 Lossy Networks", draft-ietf-roll-rpl-15 (work in 510 progress), November 2010. 512 [Levis04] Levis, P., Patel, N., Culler, D., and S. Shenker, 513 "Trickle: A Self-Regulating Algorithm for Code Propagation 514 and Maintenance in Wireless Sensor Networks"", Proceedings 515 of the First USENIX/ACM Symposium on Networked Systems 516 Design and Implementation NSDI 2004, March 2004, 517 . 519 [Levis08] Levis, P., Brewer, E., Culler, D., Gay, D., Madden, S., 520 Patel, N., Polastre, J., Shenker, S., Szewczyk, R., and A. 521 Woo, "The Emergence of a Networking Primitive in Wireless 522 Sensor Networks", Communications of the ACM, v.51 n.7, 523 July 2008, 524 . 526 [Lin08] Lin, K. and P. Levis, "Data Discovery and Dissemination 527 with DIP", Proceedings of the 7th international conference 528 on Information processing in sensor networks IPSN 2008, 529 April 2008, 530 . 532 Authors' Addresses 534 Philip Levis 535 Stanford University 536 358 Gates Hall, Stanford University 537 Stanford, CA 94305 538 USA 540 Phone: +1 650 725 9064 541 Email: pal@cs.stanford.edu 543 Thomas Heide Clausen 544 LIX, Ecole Polytechnique 546 Phone: +33 6 6058 9349 547 Email: T.Clausen@computer.org 548 Jonathan Hui 549 Arch Rock Corporation 550 501 Snd St., Suite 410 551 San Francisco, CA 94107 552 USA 554 Email: jhui@archrock.com 556 Omprakash Gnawali 557 Stanford University 558 S255 Clark Center, 318 Campus Drive 559 Stanford, CA 94305 560 USA 562 Phone: +1 650 725 6086 563 Email: gnawali@cs.stanford.edu 565 JeongGil Ko 566 Johns Hopkins University 567 3400 N. Charles St., 224 New Engineering Building 568 Baltimore, MD 21218 569 USA 571 Phone: +1 410 516 4312 572 Email: jgko@cs.jhu.edu