Transcript edinburghworkflowdec03

HPSearch and NaradaBrokering: Workflow
Scripting and Stream Management
Edinburgh December 3 2003
PTLIU Laboratory for Community Grids
Geoffrey Fox, Harshawardhan S. Gadgil, Shrideep Pallickara
Indiana University, Bloomington IN 47404
http://www.hpsearch.org
http://www.naradabrokering.org
[email protected]
Backdrop






Workflow systems have several components
Development Environment – Graphical User Interface
Specification Language such as BPEL4WS
Some interface between specification and runtime
(compiler?)
Run-time managing linkage of services, error handling
and notification
This project contributes to
• Procedural specification of workflow
• Stream management part of run-time

Workflow is sufficiently complex that we ought to agree
on a general architecture so we can each build parts
and link together
Comments on Standards




In this talk, workflow is synonymous with
“Programming the Grid/Internet”
We never agreed on a programming model for the
simpler case of “Programming a CPU” so not very
likely we will agree on standards for workflow
We did roughly agree on standards within a particular
language Fortran v. Java v C++ v C# v Lisp
We also had a little more agreement on common runtime than on languages but not complete
What to remember about this talk

All streams (flow between ports of a web service) are
handled by publish-subscribe messaging infrastructure
• Allows robust data transfer with adaptive routing e.g. allows
use of GridFTP
• Supports full concurrency inter and intra streams



Data, Streams, Files, Web Services are manipulated by
a scripting language analogous to Shell and Perl in
UNIX
http://www.hpsearch.org has the details
Software will be included in open-source release of
http://wwww.naradabrokering.org
• NB Version 0.93 today; 1.0 February 04; 2.0 for SC04
includes HPSearch
Scripting Environment I






HPSearch is designed as a scripting interface to the Internet
(Grid) using currently the Rhino implementation of Javascript
• Could use Python, Perl
• Called HPSearch because could access variables either by
URI or by search interface (To Google Web Service) but this
is not relevant here
x = ‘wsdl:http://156.56.104.155:8080/axis/services/Calculator.jws/add(5, 6)‘;
Or x = ‘wsdl:WSDL for WS/function(arguments)‘;
Returns x=11 if function adds its arguments
Could follow by y=x+1 setting y=12 etc.
Can access any data in this fashion and support normal
capabilities supported in most languages (set x=data and use
I/O)
• Perhaps prefer all I/O to go through Web Services
Scripting Environment II



So scripting environment can manipulate its own
variables and methods as usual but can also invoke any
web service with the wsdl: primitive
xpath: primitive evaluates an XPath query against a
local variable defined (say by return from a Web
service) as an XML instance
Can have multiple communicating scripting engines
WS1
WS2
Script1
WS3
WS4
WS6
WS5
Script2
This is scripting control; workflow is between web services
Script3
NaradaBrokering
Audio/Video
Conferencing Client
Computer
Modem
Minicomputer
Server
Web Service B
Peers
NaradaBrokering Broker
Network
Queues
Firewall
Stream
Workstation
Laptop computer
Peers
Web Service A
PDA
Audio/Video
Conferencing Client
Grid Messaging Substrate
SOAP+HTTP
GridFTP
RTP ….
Standard client-server
style communication.
Consumer
Service
Substrate mediated
communication removes
Consumer
transport protocol dependence.
SOAP+HTTP
GridFTP
RTP ….
Service
Messaging Substrate
Any Protocol satisfying QoS
Protocols have become overloaded e.g. MUST
use UDP for A/V latency requirements but
MUSTn’t use UDP as firewall will not support
………
NaradaBrokering

Based on a network of cooperating broker nodes
• Cluster based architecture allows system to scale in size


Originally designed to provide uniform software
multicast to support real-time collaboration linked to
publish-subscribe for asynchronous systems.
Now has several core functions
• Reliable order-preserving “Optimized” Message transport
(based on performance measurement) in heterogeneous
multi-link fashion with TCP, UDP, SSL, HTTP, and will add
GridFTP
• General publish-subscribe including JMS & JXTA and
support for RTP-based audio/video conferencing
• Distributed XML event selection using XPATH metaphor
• QoS, Security profiles for sent and received messages
• Interface with reliable storage for persistent events
Laudable Features of NaradaBrokering







Is open source http://www.naradabrokering.org
Has end-point “plug-in” as well as standalone brokers
Will have a discovery service to find nearest brokers
and manage topics
Does tunnel through many firewalls without requiring
ports to be opened
Supports JXTA, JMS (Java Message Service) and more
powerful native mode
Transit time < 1 millisecond per broker
Initial version of setup and broker network
administration module
• Currently expect to use HPSearch scripts to specify setup
NaradaBrokering Naturally Supports



Filtering of events to support different client
requirements (e.g,. PDA versus desktop, slow lines,
different A/V codecs)
Virtualization of addressing, routing, interfaces
Federation and Mediation of multiple instances of Grid
services as illustrated by
• Composition of Gridlets into full Grids (Gridlets are single
computers in P2P case)
• JXTA with peer-group forming a Gridlet



Monitoring of messages for Service management and
general autonomic functions
Fault tolerant data transport
Virtual Private Grid with fine-grain Security model
NaradaBrokering Communication




Applications interface to NaradaBrokering through
UserChannels which NB constructs as a set of links between NB
Brokers acting as “waystations” which may need to be
dynamically instantiated
UserChannels have publish/subscribe semantics with XML topics
Links implement a single conventional “data” protocol.
• Interface to add new transport protocols within the
Framework
• Administrative channel negotiates the best available
communication protocol for each link
Different links can have different underlying transport
implementations
• Implementations in the current release include support for
TCP,UDP, Multicast, SSL, RTP and HTTP.
• GridFTP most interesting new protocol
• Supports communication through proxies and firewalls such as
iPlanet, Netscape, Apache, Microsoft ISA and Checkpoint.
Manipulating Streams




flow: primitive manages streams between Web services
There is service-oriented workflow where streams are
typically implicit. Here HPSearch supports UNIX style
pipe and tee and we have trivial examples
For stream-oriented, the streams are explicit. We have
built a sophisticated system GlobalMMCS but it is
currently not supported in HPSearch
HPSearch will become control engine for
NaradaBrokering when streams are “just” message
flows on the Grid. Here one would use NB discovery
services – find streams – and monitor
• In this view a client talking to a Web Service is workflow
HPSearch Flow Example



// The input file
x = "file:///u/hgadgil/datafile.txt";
// Reverses every line in the i/p e.g. abcd becomes dcba

y1 = "156.56.104.155:5050";

// Computes the length of each line minus the last (\n or \r)

y2 = "156.56.104.155:6060";
NaradaBrokering Queue
// And finally the outputs...
z1 = "file:///u/hgadgil/reversed.txt";
y1
z1
z2 = "file:///u/hgadgil/length.txt";
x
`flow: x &> (y1 | z1), (y2 | z2)`;




T
Pipe
Pipe
y2
z2
Another Example

`flow: x &> (y1|z1 &> p,(q|storage1)), (y2|z2|storage2)`;
y1
z1
p
q
x
y2
z2
storage1
storage2
NaradaBrokering Topic (Queue)

Note this approach allows for example all workflow
streams to use RMI, GridFTP, RTP – your or rather
NaradaBrokering’s choice
Stream–oriented Workflow




As in audio-video conferencing and multimedia file
delivery where it’s the media streams that are the
“point”
Services generate and transform streams but one thinks
of streams going through services rather than services
generating streams
Multi-cast streams where video from one client sent to
all other participants in a collaborative session common
One thinks of a stream being published and
participants subscribing to it.
Pub/Sub Queue
Publish
Subscribe
XGSP Web Service MCU Architecture
Use Multiple Media servers to scale to many codecs and many
versions of audio/video mixing; should allow all e-Scientists to be connected
Session Server
XGSP-based Control
NB Scales as
distributed
Admire
Web
Services
NaradaBrokering
All Messaging
SIP
H323
Media Servers
Filters
High Performance (RTP)
and XML/SOAP and ..
Access Grid
Gateways convert to uniform XGSP Messaging
NaradaBrokering
Native XGSP
Service-oriented Workflow I

As in follow of data between different simulation
programs where one has a program (which becomes a
Web service) view and data flow between programs
often not explicitly interesting
Elastic Dislocation Inversion
Viscoelastic FEM
Viscoelastic Layered BEM
Elastic Dislocation
Pattern Recognizers
Fault Model BEM
Service-oriented Workflow II


Initial input and output files identified with perhaps a
visualization as output
In many implementations such as ours in earthquake
example one writes and reads files for stream interface
• Sometimes one wants the intermediate output files


AVS and such visualization and image processing
systems have such a model using streams
Multicast not important per-se; use a publish/subscribe
mechanism as it is fault-tolerant and higher
performance and not because of multi-cast support
Streams and Data


Scripting engine can either define topics or find them
out from NaradaBrokering discovery service
Run-time ensures that all I/O goes through
NaradaBrokering
• Note one either uses a proxy or builds NaradaBrokering
interface into Web service
• Proxy should be near Web Service as only NaradaBrokering
“guarantees” firewall penetration, fault-tolerance,
performance
• NaradaBrokering needs improved discovery system

NaradaBrokering and Scripts are distributed so no
central bottlenecks
NaradaBrokering in practice

One can “best” insert NaradaBrokering end-point
interface into each client or web service
NB Broker
Network
Client
NB
Endpt

NB Streams
WS
NB
Endpt
But proxy model easiest for existing applications
Client
Proxy
NB
Endpt
NB Broker
Network
NB Streams
Proxy
WS
NB
Endpt
“Native Communication” – cannot use added value of NB
including fault tolerance. Current GridFTP Implementation
Entities in HPSearch


Each Script is a Web Service
Each Web Service, File, Web Page has a URI and can be
accessed by a Script
• HPSearch at its heart was URI’s bound to Javascript


Publish/Subscribe system defines topics which are the URI of
streams. Note syntax is often
• topic://Session URI/stream1 with classic hierarchical
labeling
Scripts need discovery system to keep track of URI’s and in
particular the session URI (which plays role of context) -currently this is same as NaradaBrokering Discovery System
• Pub/Sub Streams typically support conversations with
related streams topic://Session URI/stream1/WS-A and
topic://Session URI/stream1/WS-B to allow Web services A
and B to interact
Publish/Subscribe Topics






One has “data” which has perhaps an intrinsic URI
For files and web pages, we have as well the location URL
I think Publish/Subscribe topic is like the URI for streams
and it is instantiated as a particular queue (or set of queues)
in NB
In NB Topics are integers (for performance), URI style or
general XML instances
Note that session topic can be thought of as “context” for
messages sent to topic as it provides intrinsic information as
to meaning of stream (cf. OGSI; WS-Addressing WSContext WS-Reliable Messaging and WS-Routing)
Topics for streams and sessions virtualize destination,
routing and context
Role of Pub/Sub Queues




One can think of N/B as providing an operating service
to transmit streams between end-points with various
value-added capabilities
Messages are the units of a stream
Events are messages with time-stamps (which could be
absent); so events are messages and vice versa
Streams are ordered collections of messages
• NB manipulates streams and collections of streams
• Delivery is guaranteed order preserving

NB provides a virtual stream desktop which you can
use to manipulate streams in same way you manipulate
files in conventional O/S
Multiple Input and Output Ports


We can deal with Web Services with multiple input and
output using an array notation but the &> Tee and |
Pipe notation get clumsy
So can use explicit notation such as
• x.port[0].publish = NBTopicA;
• y1.port[0].subscribe = NBTopicA;
• y2.port[0].subscribe = NBTopicA;


This would also be natural way of implementing
stream-oriented workflow
Errors and notifications would be easy in this syntax
• notifyTOPIC = SessionTOPIC + ‘/notify’;
• x.notify.publish = notifyTOPIC;
• scriptasaWS.port[1].subscribe = notifyTOPIC;
HPSearch Administrative Interface to NB

One can build administrative policies and procedures
by flowing administrative and monitoring data to
appropriate scripting engines
•
•
•
•
•
•


performanceTOPIC = SessionTOPIC + ‘/performance’;
nbws = NBDiscover(“aggregateperformancews”)
nbws.performancedata.publish = performanceTOPIC;
scriptasaWS.port[2].subscribe = performanceTOPIC;
Niftyperformanceanalyser(scriptasaWS.port[2]);
…….
This example pipes performance data from
NaradaBrokering and spawns some analysis
NB provides for each link (broker to broker, broker to
end-point) available bandwidth, used bandwidth,
latency etc.
Other NB Features to be added to HPSearch







Full details of available Brokers and Stable storage
Pending queue sizes
Message statistics – size, number per second, time since
since last message – at brokers and end-points
Current stream sequence number at different parts of
pipeline from source to destination
Heartbeat Information
Active Topics and list of publishers and subscribers
(subject to security restrictions)
Fault tolerance statistics including those subscribed
end-points which are “down”
Transit Delay (Milliseconds)
Mean transit delay for message samples in
NaradaBrokering: Different communication hops
9
8
7
6
5
4
3
2
1
0
hop-2
hop-3
hop-5
hop-7
100
1000
Message Payload Size (Bytes)
Pentium-3, 1GHz,
256 MB RAM
100 Mbps LAN
JRE 1.3 Linux
Standard Deviation for message samples in NaradaBrokering
Different communication hops - Internal Machines
0.8
hop-2
hop-3
hop-5
hop-7
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1000
1500
2000
2500
3000
3500
Message Payload Size
(Bytes)
4000
4500
5000
Average delays per packet for 50 video-clients
NaradaBrokering Avg=2.23 ms, JMF Avg=3.08 ms
60
NaradaBrokering-RTP
JMF-RTP
Delay (Milliseconds)
50
40
30
20
10
0
0
200 400 600 800 1000 1200 1400 1600 1800 2000
Packet Number
Average jitter (std. dev) for 50 video clients.
NaradaBrokering Avg=0.95 ms, JMF Avg=1.10 ms
8
NaradaBrokering-RTP
JMF-RTP
Jitter (Milliseconds)
7
6
5
4
3
2
1
0
0
200 400 600 800 1000 1200 1400 1600 1800 2000
Packet Number