1. Ergodic Continues Hidden Markov
Models for work load
characterization
•
• By
G RAGHU(14IT05F)
NITYA PRIYA(14IT14F)
2. •
● Performance evolution of computer system requires to
test different alternatives under given workload.
●
What is work load ?
3. Workload
• The workload of a system can be defined as the
set of all inputs that the system receives from its
environment during any given period of time.
•
HTTP
requests
Web Server
Adapted from Menascé & Almeida. 3
4. • Generally real time computing environment is not
repeatable.
● Workload characterization using a model play a
fundamental role in many computer architecture areas.
e.g.
• understand the key resource usage application,
q
• to guide the selection of the programs for bench mark
tests.
5. • workload modelling starts with measured data about the
workload.
• Data is recorded as a log file or trace of the workload –
related events that happened in a particular system.
•
• Generally two ways to evaluate the performance of a
computer.
•
(i) use direct traced work load for analysis.
(ii) create a model.
●
6. Our approach is based on the idea to treat the sequences of
virtual pages produced by the running application as time
varying discrete-time series of data and to analyse with
statistical techniques.
In other words we consider the similar kind of process
obtained by the same type of work load and re-estimating
with the hidden markov model.
7. Our estimated model can be used in two ways:
i) To determine the which workload belongs to the current running
application.
i)
ii) To generate log file.
e.g. running process coming from c compiler or perl interpreter or
from the chess game and so on.
This knowledge can be used ,for example to better to manage the
requested resources.
8. Hmm for workloads characterization:
parameters:
a) Since the page references are time varying ,we used
short-time spectral analysis,
b) Sequence of virtual pages divided into short sections and those are analysed
by DFT.
c) As in proposed approach issue is related to comparison the log-spectral data,
define the distance between two log sceptical data.
d)
9. Spectral distance between the two log spectra is simply Euclidian distance
between the two spectral sequences
Hidden Markov Model:
The basic Markov model is the Markov chain, which is represented with a graph
composed by a set of N states; the graph describes the fact that the probability of
the next event depends on the previous event.
Markov models are too simple to describe complex real
systems.so we will go with hidden markov model.
10. Workload classification:
For dynamic characterization of
processes, the address field of the traces has been extracted. In this
way we have obtained a sequence of virtual addresses generated by
the processor during execution. For converting the trace of
addresses into trace of virtual pages, the sequence of addresses has
been divided by the page dimension, which we set to 4096 bytes.
Once the sequence of virtual pages has been obtained from
every trace and thus for every process, we have first tried
to use discrete HMMs for their classification.
11. Single trace classification;
The sequences are floating point
sequences. If we want to use a discrete HMM for analyzing the
cepstral data, the continuous data should be turned in a discrete
sequence. We did this operation using vector quantization.
Quantitization: quantization technique from signal
processing which allows the modeling of probability density
functions.
12. Program behavior modeling:
Single Trace Classification of the traces,
taking as parameter the virtual pages, has obtained satisfactory
results. Each trace has been obtained running a program with
different inputs.
In this modeling using several traces of the same workload for
classifying the program behavior using discreet and continuous
HMM.
15. CONCLUSION:
In this paper we describe an
approach for workload characterization using ergodic
hidden Markov models.
classification accuracy rate about 76%.