Professional Documents
Culture Documents
Malleable-Lab: A Tool For Evaluating Adaptive Online Schedulers On Malleable Jobs
Malleable-Lab: A Tool For Evaluating Adaptive Online Schedulers On Malleable Jobs
I. I NTRODUCTION
The emergence of multi-core computers has led to explosive
development of parallel applications and hence the need of efficient schedulers for parallel jobs. Adaptive online schedulers
have recently been proposed to exploit the multiple processor
resource [1], [2], [3] and they have shown good promise in
theory. To verify the effectiveness of these parallel schedulers,
it will be reassuring to test them extensively with various
parallel workloads. Unfortunately it is still unknown how the
job mixes will eventually evolve for multi-core computers;
moreover, it is also non-obvious how the parallelism of a
typical job will look like. An ideal parallel workload model for
schedulers should thus allow the user to vary the parallelism
profiles of individual jobs as well as the job arrival patterns.
With the adaptive schedulers, there is a special need to test
the responsiveness and stability of a given request-allocate
cycle. This need arises because, with adaptive schedulers, the
1066-6192/10 $26.00 2010 IEEE
DOI 10.1109/PDP.2010.14
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.
100
800
Step
80
700
Log
Poly(II)
Poly(I)
Parallelism
Parallelism
600
60
Exp
40
Ramp
Impulse
20
Impulse
Exp
Ramp
Poly(I)
500
400
300
200
Step
Poly(II)
Log
100
Time
Time
(a)
(b)
10
Fig. 1. (a) Seven different types of parallelism variation curves specified by Step, Log, Poly(II), Ramp, Poly(I), Exp, Impulse functions. (b) Seven parallelism
variation curves with each one having the same work and length, hence the the same average parallelism.
12
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.
Fig. 2.
Sample parallel program segments and their corresponding parallelism variation over time.
OF
A DAPTIVE S CHEDULING
13
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.
300
140
Parallelism
AGDEQ
ABGDEQ
100
80
60
40
200
150
100
50
20
0
0
Parallelism
AGDEQ
ABGDEQ
250
Parallelism or Request
Parallelism or Request
120
10
0
0
15
Time
(a)
300
Parallelism
AGDEQ
ABGDEQ
200
150
100
Parallelism
AGDEQ
ABGDEQ
250
Parallelism or Request
Parallelism or Request
250
200
150
100
50
50
10
0
0
15
Time
10
15
Time
(c)
(d)
600
800
Parallelism
AGDEQ
ABGDEQ
400
300
200
Parallelism
AGDEQ
ABGDEQ
700
Parallelism or Request
500
Parallelism or Request
15
(b)
300
0
0
10
Time
600
500
400
300
200
100
0
0
100
10
0
0
15
Time
10
15
Time
(e)
(f)
Fig. 3. Transient responses of ABG-DEQ and AG-DEQ with respect to (a) Step, (b) Log, (c) Poly(II), (d) Ramp, (e) Poly(I) and (f) Exp parallelism profile.
14
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.
15
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.
5
4.5
AGDEQ
ABGDEQ
ResponseTime Ratio
EQUI
AGDEQ
ABGDEQ
3.5
3
2.5
6
5
4
3
2
2
1.5
1
1
0
Step
Log
0.5
Exp Impulse
(a)
1.5
Load
2.5
(b)
Fig. 4. Response time of ABG-DEQ, AG-DEQ and EQUI on (a) seven different parallelism profiles in ideal, unconstrained environment and (b) a range of
workloads with mixed parallelism profiles.
1.2
AGDEQ
ABGDEQ
0.9
0.8
0.7
0.8
Utilization
Utilization
0.6
0.6
0.5
0.4
0.4
0.2
0.3
0.2
0
Step
Log
Exp Impulse
(a)
EQUI
AGDEQ
ABGDEQ
0.5
1.5
Load
2.5
(b)
Fig. 5. Utilization of ABG-DEQ, AG-DEQ and EQUI on (a) seven different parallelism profiles in ideal, unconstrained environment and (b) a range of
workloads with mixed parallelism profiles.
16
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.
1.6
Response Time Ratio (EQUI / AGDEQ)
1.6
1.4
1.2
=0
=0.02
=0.04
=0.06
=0.08
=0.1
0.8
0.6
0.4
0
0.5
1.5
Load
2.5
1.4
1.2
0.6
0.4
0
=0
=0.02
=0.04
=0.06
=0.08
=0.1
0.8
0.5
(a)
2.5
3.5
=0
=0.02
=0.04
=0.06
=0.08
=0.1
2.5
1.5
0.5
1.5
Load
2.5
=0
=0.02
=0.04
=0.06
=0.08
=0.1
2.5
1.5
1
0
(c)
Fig. 6.
(b)
3.5
1
0
1.5
Load
0.5
1.5
Load
2.5
(d)
Impacts of overhead on ABG-DEQ, AG-DEQ and EQUI with respect to response time and utilization
17
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.
18
Authorized licensed use limited to: UNIVERSIDADE FEDERAL DO RIO GRANDE DO NORTE. Downloaded on June 07,2010 at 10:56:04 UTC from IEEE Xplore. Restrictions apply.