-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathindex.html
435 lines (408 loc) · 21.6 KB
/
index.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Varys - Application-Aware Network Scheduling</title>
<link rel="stylesheet" href="css/bootstrap.min.css">
<link rel="stylesheet" href="font-awesome-4.1.0/css/font-awesome.min.css">
<link rel="stylesheet" href="css/custom.css">
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','//www.google-analytics.com/analytics.js','ga');
ga('create', 'UA-4096845-3', 'varys.net');
ga('send', 'pageview');
</script>
<!-- HTML5 shim and Respond.js IE8 support of HTML5 elements and media queries -->
<!--[if lt IE 9]>
<script src="https://oss.maxcdn.com/libs/html5shiv/3.7.0/html5shiv.js"></script>
<script src="https://oss.maxcdn.com/libs/respond.js/1.4.2/respond.min.js"></script>
<![endif]-->
</head>
<body data-spy="scroll" data-offset="70" data-target="#myNavbar" id="home">
<div class="navbar navbar-default navbar-fixed-top" role="navigation">
<div class="container">
<div class="navbar-header">
<button type="button" class="navbar-toggle" data-toggle="collapse" data-target=".navbar-responsive-collapse">
<span class="icon-bar"></span>
<span class="icon-bar"></span>
<span class="icon-bar"></span>
</button>
<a class="navbar-brand" href="#home">Varys</a>
</div>
<div class="navbar-collapse collapse navbar-responsive-collapse" id="myNavbar">
<ul class="nav navbar-nav">
<li class="active"><a href="#overview">Overview</a></li>
<li><a href="#performance">Performance</a></li>
<li><a href="#research">Research</a></li>
<li><a href="#faq">FAQ</a></li>
</ul>
</div><!--/.nav-collapse -->
</div>
</div>
<div class="jumbotron">
<div class="container" align="center">
<h1>Varys</h1>
<p>Application-Aware Network Scheduler</p>
<p> </p>
<p align="center"><a class="btn btn-success" role="button" href="http://www.github.com/coflow/varys"><i class="fa fa-github fa-lg"></i> Download Varys</a></p>
</div>
</div>
<div class="container">
<div class="row row-padded">
<div class="col-lg-3">
<h3><i class="fa fa-rocket fa-lg"></i> Speed</h3>
<p class="lead text-muted">3x faster communication.</p>
<p>Using a novel coflow scheduling algorithm, Varys minimizes communication times of data-parallel applications without starving anyone. </p>
</div>
<div class="col-lg-3">
<h3><i class="fa fa-clock-o fa-lg"></i> Predictability</h3>
<p class="lead text-muted">2x more often on time.</p>
<p>Varys uses admission control to support soft deadlines for timely completion of most coflows. </p>
</div>
<div class="col-lg-3">
<h3><i class="fa fa-gears fa-lg"></i> Multi-Tenancy</h3>
<p class="lead text-muted">Informed cluster sharing.</p>
<p>Using Varys, coexisting applications, frameworks, or users avoid collateral damage when optimizing respective communication performance. </p>
</div>
<div class="col-lg-3">
<h3><i class="fa fa-magic fa-lg"></i> Transparency</h3>
<p class="lead text-muted">No changes to user jobs.</p>
<p>Transfers data from sources to destinations without tweaking the number of connections or playing with bytes-in-flight in each connection. </p>
</div>
</div>
</div>
<hr>
<div class="container">
<div class="row row-padded">
<div class="col-lg-12">
<h1 id="overview">What is Varys?</h1>
<p>
Varys is an open source network manager/scheduler that aims to improve communication performance of Big Data applications.
Its target applications/jobs include those written in Spark, Hadoop, YARN, BSP, and similar data-parallel frameworks.
</p>
<p>
Varys provides a simple API that allows data-parallel frameworks to express their communication requirements as <a href="#coflow">coflows</a> with minimal changes to the framework.
Using coflows as the basic abstraction of network scheduling, Varys implements <a href="#scheduler">novel schedulers</a> either to make applications faster or to make time-restricted applications complete within deadlines.
</p>
</div>
</div>
<div class="row row-padded">
<div class="col-lg-4">
<h3>How to Use it?</h3>
<p>
Enabling Varys in <em>all</em> the applications written in a framework can be as simple as changing the <samp>send</samp> and <samp>recv</samp> calls (or their equivalents) to their analogous <samp>put</samp> and <samp>get</samp> methods in the coflow API, which are exposed through the <samp>VarysClient</samp> class.
</p>
<p>
Existing user-written jobs can take advantage of coflows <em>without</em> any modifications.
</p>
<p>
<!-- <a href="http://www.github.com/coflow/varys-demo">See</a> the coflow API in action in a simplistic data-parallel framework. -->
</p>
</div>
<div class="col-lg-8">
<h3>The Coflow API</h3>
<table class="table table-striped table-bordered table-condensed small">
<tbody>
<tr>
<td>Method Signatures in Scala</td>
<td>Calling Process</td>
</tr>
<tr>
<td><samp>def registerCoflow(coflowDesc: CoflowDescription): String</samp></td>
<td>Driver</td>
</tr>
<tr>
<td><samp>def putObject[T](objectId: String, obj: T, coflowId: String, size:Long)</samp></td>
<td>Sender</td>
</tr>
<tr>
<td><samp>def getObject[T](objectId: String, coflowId: String): T</samp></td>
<td>Receiver</td>
</tr>
<tr>
<td><samp>def unregisterCoflow(coflowId: String)</samp></td>
<td>Driver</td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
<hr>
<div class="container">
<div class="row">
<div class="col-lg-12">
<h1 id="performance">Performance</h1>
<p>
Varys is most effective when applied to a number of concurrently-running coflows from different jobs.
We have evaluated Varys by deploying it on a 100-node Amazon EC2 cluster and running traces collected from a 3000-node production Hive/Hadoop cluster at Facebook, and we compare Varys's performance with that of TCP fair sharing without using the coflow abstraction.
</p>
</div>
</div>
<div class="row">
<div class="col-lg-6" align="center">
<h3>Minimizes Communication Time</h3>
<img src="images/mincct.png" class="img-responsive img-thumbnail" alt="As jobs spend more time in communication, Varys improves their communication performance even more." title="As jobs spend more time in communication, Varys improves their communication performance even more.">
</div>
<div class="col-lg-6" align="center">
<h3>Meets More Deadlines</h3>
<img src="images/deadline.png" class="img-responsive img-thumbnail" alt="As deadlines become less stringent, Varys completes more and more coflows within time." title="As deadlines become less stringent, Varys completes more and more coflows within time.">
</div>
</div>
<div class="row">
<div class="col-lg-12">
<br/>
<p>
More details on different aspects of Varys's performance—comparisons with newer per-flow prioritization schemes or behavior with increasing load—can be found in accompanying research <a href="#papers">papers</a>.
</p>
</div>
</div>
</div>
<hr>
<div class="container">
<div class="row">
<div class="col-lg-12">
<h1 id="research">Research</h1>
<h3 id="coflow">The Coflow Abstraction</h3>
<p>
Despite the differences of end goals and/or programmability among data-parallel frameworks, their communication is structured, takes place between groups of machines in successive computation stages, and often involves a collection of parallel flows.
More importantly, communication in these applications exhibit the <em>all-or-nothing</em> property—<em>a communication stage cannot finish until all its flows have completed</em>.
</p>
<p>
The coflow abstraction represents such collections of parallel flows to convey job-specific communication requirements – for example, minimizing completion time or meeting a deadline – to the network and enables application-aware network scheduling.
Currently, we assume that the amount of data each flow needs to transfer is known before it starts.
The flows of a coflow are <em>independent</em> in that the input of a flow does not depend on the output of another in the same coflow, and the endpoints of these flows can be in one or more machines.
Effectively, one can consider a coflow to be a <em>data-parallel network job</em>.
</p>
</div>
</div>
<div class="row">
<div class="col-lg-5">
<p>
Examples of coflows include the shuffle between the mappers and the reducers in MapReduce and the communication stage in the bulk-synchronous parallel (BSP) model.
In fact, coflows can express most communication patterns between successive computation stages of data-parallel applications.
</p>
<p>
Finally, traditional point-to-point communication is also a coflow with a single flow.
</p>
</div>
<div class="col-lg-7">
<table class="table table-striped table-bordered table-condensed small">
<tbody>
<tr>
<td></td>
<td>Example Frameworks</td>
<td>Coflow Structure</td>
</tr>
<tr>
<td>Shuffle in dataflow pipelines</td>
<td>Spark, Hadoop</td>
<td>Many-to-Many</td>
</tr>
<tr>
<td>Broadcast</td>
<td>Spark</td>
<td>One-to-All</td>
</tr>
<tr>
<td>Aggregation</td>
<td>Spark, Hadoop</td>
<td>Many-to-One</td>
</tr>
<tr>
<td>Global communication barriers</td>
<td>BSP</td>
<td>All-to-All</td>
</tr>
<tr>
<td>Parallel read/write from/to distributed storage</td>
<td>HDFS</td>
<td>Many One-to-One</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="row">
<div class="col-lg-8">
<h3 id="scheduler">How/Why Does Coflow Scheduling Work in Varys?</h3>
<p>
Varys schedules coflows using heuristics developed to solve the recently discovered <em>concurrent open shop scheduling with coupled resources</em> problem.
Similar to most scheduling problems, optimal coflow scheduling is computationally intractable.
</p>
<p>
From 10000 feet, the algorithm proceeds in two steps: Varys keeps an ordered list of coflows, preempting whenever necessary. Next, it uses an algorithm to allocate minimum required bandwidth to each flow of a coflow to complete all of its flows together.
</p>
<p>
In essence, Varys minimizes interleaving across coflows to minimize coflow completion times (CCT), while sacrificing flow-level fairness or flow completion times (FCT) that are not correlated to a job's completion time.
</p>
<h4>Example</h4>
<p>
Consider two coflows C<sub>1</sub> and C<sub>2</sub>; C<sub>1</sub> has three flows transferring 1, 2, and 4 units of data, while C<sub>2</sub> has two flows transferring 2 data units each.
</p>
<p>
Per-flow schemes focus on fair sharing or decreasing FCT, without caring about application's completion. Varys does the opposite.
</p>
</div>
<div class="col-lg-4">
<h3> </h3>
<div class="thumbnail">
<img src="images/ex1-input.png" class="img-responsive">
<div class="caption">
<p>
Coflow scheduling over a 3x3 datacenter fabric with three ingress/egress ports. Flows in ingress ports are organized by their destinations and color-coded by coflows: C<sub>1</sub> in orange and C<sub>2 </sub> in blue.
</p>
</div>
</div>
</div>
</div>
<div class="row">
<div class="col-lg-4">
<div class="thumbnail">
<img src="images/ex1-fair.png" class="img-responsive">
<div class="caption">
<p>
<em>Per-flow fairness:</em> The average FCT and CCT are respectively 3.4 and 4 time units.
</p>
</div>
</div>
</div>
<div class="col-lg-4">
<div class="thumbnail">
<img src="images/ex1-pdq.png" class="img-responsive">
<div class="caption">
<p>
<em>Per-flow prioritization:</em> The average FCT and CCT are respectively 2.8 and 3.5 time units.
</p>
</div>
</div>
</div>
<div class="col-lg-4">
<div class="thumbnail">
<img src="images/ex1-varys.png" class="img-responsive">
<div class="caption">
<p>
<em>Varys with inter-coflow scheduling:</em> The average FCT and CCT are both 3 time units.
</p>
</div>
</div>
</div>
</div>
<div class="row">
<div class="col-lg-12">
<h3 id="papers">Relevant Publications</h3>
<p>More details on Varys, its performance, the coflow abstraction, and Varys's predecessor system (Orchestra) can be found in the following papers. </p>
<ul>
<li><a href="http://www.mosharaf.com/wp-content/uploads/varys-sigcomm14.pdf">Efficient Coflow Scheduling with Varys</a>, Mosharaf Chowdhury, Yuan Zhong, Ion Stoica, ACM SIGCOMM, 2014.</li>
<li><a href="http://www.mosharaf.com/wp-content/uploads/coflow-hotnets2012.pdf">Coflow: A Networking Abstraction for Cluster Applications</a>, Mosharaf Chowdhury, Ion Stoica, ACM HotNets-XI, 2012.</li>
<li><a href="http://www.mosharaf.com/wp-content/uploads/orchestra-sigcomm11.pdf">Managing Data Transfers in Computer Clusters with Orchestra</a>, Mosharaf Chowdhury, Matei Zaharia, Justin Ma, Michael Jordan, Ion Stoica, ACM SIGCOMM, 2011.</li>
</ul>
<p>Varys also exposes APIs to find the machines with the most and the least network usage – <samp>getBestRxMachines</samp> and <samp>getBestTxMachines</samp> – that can be used to perform, for example, network-aware replica placement in HDFS as outlined in the following paper. </p>
<ul>
<li><a href="http://www.mosharaf.com/wp-content/uploads/sinbad-sigcomm13.pdf">Leveraging Endpoint Flexibility in Data-Intensive Clusters</a>, Mosharaf Chowdhury, Srikanth Kandula, Ion Stoica, ACM SIGCOMM, 2013.</li>
</ul>
</div>
</div>
</div>
<hr>
<div class="container">
<div class="row">
<div class="col-lg-12">
<h1 id="faq">Varys FAQ</h1>
<p class="question">Is Varys a new networking protocol? Does it depend on a new one?</p>
<p class="answer">
No.
Varys runs in the application layer on top of underlying transport protocols (e.g., TCP).
No changes to your hardware or OS is required.
Except for minor changes to the frameworks, user jobs can transparently take advantage of Varys as well.
</p>
<p class="question">Varys seems to have <samp>put</samp> and <samp>get</samp> methods. Is Varys a storage system? Does it implement one underneath?</p>
<p class="answer">
No, Varys does not have/need its own storage system.
It sends and receives from user process (if data is in-memory) or from disk (if data is already stored there).
</p>
<p class="question">Will it always improve performance?</p>
<p class="answer">
Varys will improve performance most of the time, specially when you have a multi-user, multi-framework cluster.
Currently, overheads due to Varys's centralized design sometimes make it less useful for sub-second jobs and cause deadline misses on rare occasions.
We are working on both directions: to make it more efficient by design and to better engineer the current system.
</p>
<p class="question">Which frameworks support Varys?</p>
<p class="answer">
We are working on ensuring Varys support in Spark and plan on enabling Hadoop MapReduce 2.0 as well.
We will appreciate your help in any of these endeavors.
</p>
<p class="question">Which languages does Varys support?</p>
<p class="answer">
Currently Varys provides an interface only to JVM-based languages.
It can support non-JVM languages as well, but we need your help to implement the language-specific interface to Varys.
</p>
<p class="question">What license is Varys under?</p>
<p class="answer">
Varys is under the <a href="http://www.apache.org/licenses/LICENSE-2.0.html">Apache 2.0 license</a>.
</p>
<p class="question">How can I contribute?</p>
<p class="answer">
If you would like to report a bug, please <a href="https://github.com/coflow/varys/issues">create a new issue</a>.
</p>
<p class="answer">
If you have already resolved one, do send us a pull request on <a href="https://github.com/coflow/varys">GitHub</a>.
We want to hear about your experience using Varys.
</p>
<p class="answer">
Finally, if you want to discuss a new idea, talk to us at the <a href="mailto:[email protected]">Varys Developers</a> mailing list.
</p>
<p class="question">Where can I get more help?</p>
<p class="answer">
Please post on the <a href="mailto:[email protected]">Varys Users</a> mailing list. We will be glad to help!
</p>
</div>
</div>
</div>
<hr>
<div class="container">
<p class="text-muted">
Varys is being developed in the <a href="https://amplab.cs.berkeley.edu">AMPLab</a> by <a href="http://www.mosharaf.com">Mosharaf Chowdhury</a>, <a href="http://www.columbia.edu/~yz2561/">Yuan Zhong</a>, and <a href="http://www.cs.berkeley.edu/~istoica">Ion Stoica</a>.
</p>
<p class="text-muted">
This research is supported in part by NSF CISE Expeditions Award CCF-1139158, LBNL Award 7076018, and DARPA XData Award FA8750-12-2-0331, and gifts from Amazon Web Services, Google, SAP, The Thomas and Stacey Siebel Foundation, Apple, Inc., C3Energy, Cisco, Cloudera, EMC, Ericsson, Facebook, GameOnTalis, Guavus, HP, Huawei, Intel, Microsoft, NetApp, Pivotal, Splunk, Virdata, VMware, WANdisco and Yahoo!.
</p>
</div>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/1.11.0/jquery.min.js"></script>
<script src="http://netdna.bootstrapcdn.com/bootstrap/3.1.1/js/bootstrap.min.js"></script>
<script>
/**
* Check a href for an anchor. If exists, and in document, scroll to it.
* If href argument ommited, assumes context (this) is HTML Element,
* which will be the case when invoked by jQuery after an event
*/
function scroll_if_anchor(href) {
href = typeof(href) == "string" ? href : $(this).attr("href");
// You could easily calculate this dynamically if you prefer
var fromTop = 60;
// If our Href points to a valid, non-empty anchor, and is on the same page (e.g. #foo)
// Legacy jQuery and IE7 may have issues: http://stackoverflow.com/q/1593174
if(href.indexOf("#") == 0) {
var $target = $(href);
// Older browser without pushState might flicker here, as they momentarily
// jump to the wrong position (IE < 10)
if($target.length) {
$('html, body').animate({ scrollTop: $target.offset().top - fromTop });
if(history && "pushState" in history) {
history.pushState({}, document.title, window.location.pathname + href);
return false;
}
}
}
}
// When our page loads, check to see if it contains and anchor
scroll_if_anchor(window.location.hash);
// Intercept all anchor clicks
$("body").on("click", "a", scroll_if_anchor);
</script>
</body>
</html>