A Prediction Service For Remos and QuO

1
A Prediction ServiceFor Remos and QuO

• Peter A. Dinda
• CMU SCS

2
Outline

• Bette Davis
• Remos in QuO
• Peters Prediction Service in Remos
• Software abstractions
• Implementation
• Example

3
Bette Davis Hello World Program
Choose whether to fetch a small or large image
from the Bette Davis Server based on predicted
network and host load 3 Systems (BBN QuO, CMU
Remos, Peters Prediction Service) 10 Programming
Languages (i386 assembly, C, C, Java, Fortran
77, CORBA IDL, BBN SDL, BBN CDL, Perl, Sh)
Network and Host Load Predictions
4
Abstract Model of QuO
Client
Contract
ORB and Other Magic
Contract
Server
5
Remos in QuO
Client
Contract
...
ORB and Other Magic
Impedence Mismatch
Remos Modeler (Query Interface)
Remos Collectors (LAN SNMP, WAN Bench, Host Load)
Server
6
Prediction in Remos
Client
Contract
...
ORB and Other Magic
Remos Modeler (Query Interface)
Remos Modeler (Query Interface)
Impedence Mismatch
Past
Present
Future
Prediction
MIB
Server
Collectors (LAN SNMP, WAN Bench, Host Load)
7
What is this Impedence Mismatch?
Queries
Prediction Pipeline (Inherently Stream Oriented)
Measurement Stream
8
PPL Peters Prediction Library
Abstractions
Implementations
Data Sequence
Best Mean, AR, MA, ARMA, ARIMA, ARFIMA
Modeler
Best Mean, AR, MA, ARMA, ARIMA, ARFIMA
Model
Data Stream
Prediction Stream
Predictor
EtaTheta
Data path
Evaluator
Creates
Statistical Metrics
9
A Prediction Service For Remos 1/2
Trigger Refit
Buffer
Statistical Metrics
Modeler
zt1,t1m
zt2,t2m
Evaluator
zt,tm
...
Model
...
...
...
zt1,t3
zt2,t4
zt,t2
...
zt1,t2
zt2,t3
zt,t1
...
Predictor
Predictor
zt
zt1
zt2
zt
...
zt1
zt2
...
Output Measurement And Prediction Stream
Measurement Device (eg, Collector)
10
A Prediction System For Remos 2/2
Remos Modeler Present and Near Past Queries
Remos Modeler Future Queries
Buffer
Buffer
Output Measurement And Prediction Stream
11
Implementation Goals

• Completely flexible component placement
• Flexible communication
• No CORBA/Java/Whatever dependence
• Support most internet protocols (UDP, TCP,
  Multicast, pipes, files, etc)
• Portability
• No CORBA, no threads, no Java, minimal Unix
  assumptions
• Lessons learned after initial CORBA, LDOS
  implementation attempts

12
The Mirror C Template Class 1/2
UDP Streams
UDP Streams
TCP Streams
TCP Streams
File Streams
File Streams
Data Forwarder
Pipes
Pipes
Timed Callbacks
Callbacks
TCP Stream Connection Reqs
TCP UDP
TCP Req/Resp Connection Reqs
Server
13
The Mirror C Template Class 2/2

• Parameterized by handlers
• Default handlers for 2 kinds of Mirrors
• Raw Data
• Serializeable Objects
• And several behaviors
• Buffering, Req/Resp, Compute-on-input

14
Example Measurement Buffer 1/2

• template ltclass SOURCEHANDLER,class
  SOURCECONNECTHANDLER,class REQUESTRESPONSEHANDLER,
  class REQUESTRESPONSECONNECTHANDLERgt class
  Mirror public AbstractMirror
• template ltclass SERIN, class COMPUTE, class
  SEROUTgt class GenericSerializeableInputComputeOutp
  utMirror public SerializeableMirrorltSERIN,
  SerializeableMirrorInputComputeOutputHandlerltSERIN
  ,COMPUTE,SEROUTgt, GenericMirrorNewConnectionHandle
  r,NullHandler, NullHandlergt
• template ltclass SERINgt class BufferingSerializeabl
  eMirror public SerializeableMirrorltSERIN,Serial
  izeableMirrorInputHandlerltSERINgt,GenericMirrorNewC
  onnectionHandler,
  BufferDataRequestResponseHandlerltSERINgt,
  GenericMirrorNewRequestResponseConnectionHandlergt
  
• typedef BufferingSerializeableMirrorltMeasurementgt
  MeasureBuffer

15
Example Measurement Buffer 1/2

• MeasureBuffer mirror(depth)
• for (i2iltargci)
• ep new EndPoint
• if (ep-gtParse(argvi)) exit(-1)
• mirror.AddEndPoint(ep)
• mirror.Run()

Server
ReferenceltBufferDataRequest,BufferDataReplyltMeasur
ementgtgt ref BufferDataRequest req BufferDataRepl
yltMeasurementgt repl ref.ConnectTo((new
EndPoint())-gtParse(argv2)) req.numnum ref.Cal
l(req,repl) for (i0iltrepl.numi)
repl.datai.Print(stderr)
Client
16
Command Lines From Hell
loadserver 1000000 targetfilestdout
connecttcp5000 targetudp239.99.99.995000
load2measure 0 sourcefilestdin
connecttcp5001 targetfilestdout
measurebuffer 100 sourcefilestdin
targetfilestdout servertcp5002 predserver
8 sourcefilestdin targetfilestdout
connecttcp5003 targetudp239.99.99.995003
predbuffer 100 sourcefilestdin
targetfilestdout servertcp5004 predclient
sourcefilestdin
Measure Load and streams to stdout, multicast,
connected TCP sessions
Accept stream from stdin, make generic, stream to
TCP and stdout
Buffer stdin stream, stream to stdout, accept tcp
requests for buffer contents
Predict stdin stream 8 steps ahead, stream preds
to stdout, TCP, and multicast
Buffer predictions from stdin, stream to stdout,
accept tcp requests for buffer contents
Accept stdin pred stream and print
17
Meeting Implementation Goals

• Completely flexible component placement
• Mostly, yes
• Open implementation issues Instantiation and
  naming problem Who instantiates components, how
  do we find them, etc
• This is really a Remos/QuO issue - we all need to
  solve it
• Database?
• Flexible communication
• Yes. You name it, you can communicate with it on
  as many channels as you want
• Portability
• Yes. You only need g 2.7.2 or later or similar
  C compiler
• Running on Linux, FreeBSD, Solaris, DUX

18
Zukunft

• Finish integration with Remos
• Partially me
• Finish integration of Remos with QuO
• NOT ME
• Address instantiation and naming problem
• Database? Globus?
• NOT ME
• Use prediction service to dynamically map a RT
  interactive application and show benefits
• DV, OpenMap, DynBench, Feature tracking...
• ME ME ME

19
Vergangenheit

• Host load study (LCR98, SciProg99)
• Self-similarity, epochal behavior, etc.
• Study of linear models for host load prediction
  (TR 98-174, sub. to HPDC99)
• Statistical statement AR(16) models or better
  are appropriate for Host Load Pred
• Initial flailing on applications (WPDRTS99)