(Paper) Task scheduling algorithm for multicore processor system for minimizing recovery time in case of single node fault

Task scheduling algorithm for multicore processor
system for minimizing recovery time in case of
single node fault
Shohei Gotoda, Minoru Ito
Nara Institute of Science and Technology
Nara, Japan
{shohei-g, ito}@is.naist.jp
Naoki Shibata
Shiga University
Shiga, Japan
shibata@biwako.shiga-u.ac.jp
Abstract—In this paper, we propose a task scheduling al-
gorithm for a multicore processor system which reduces the
recovery time in case of a single fail-stop failure of a multicore
processor. Many of the recently developed processors have
multiple cores on a single die, so that one failure of a computing
node results in failure of many processors. In the case of a failure
of a multicore processor, all tasks which have been executed
on the failed multicore processor have to be recovered at once.
The proposed algorithm is based on an existing checkpointing
technique, and we assume that the state is saved when nodes
send results to the next node. If a series of computations that
depends on former results is executed on a single die, we need
to execute all parts of the series of computations again in
the case of failure of the processor. The proposed scheduling
algorithm tries not to concentrate tasks to processors on a die.
We designed our algorithm as a parallel algorithm that achieves
O(n) speedup where n is the number of processors. We evaluated
our method using simulations and experiments with four PCs.
We compared our method with existing scheduling method, and
in the simulation, the execution time including recovery time in
the case of a node failure is reduced by up to 50% while the
overhead in the case of no failure was a few percent in typical
scenarios.
I. INTRODUCTION
Recently, cloud computing is growing popular, and various
kinds of services are being delivered by cloud computing. In
order to make these services more reliable and to improve
availability, we need mechanisms to prepare for failure of com-
puting nodes. On the other hand, almost all high-performance
processors are now designed as multicore processors. A mul-
ticore processor is a semiconductor chip(a die, hereafter) on
which multiple processor cores are implemented, with each
processor core capable of executing an independent task. In
case of failure of a memory or of other devices shared among
all processor cores, all processors on a die simultaneously
stop execution of their tasks, in which case all of these tasks
need to be recovered. Since communication between processor
cores on a die has a significantly larger bandwidth and smaller
latency than Ethernet or other networking technologies for
connecting computers, existing task schedulers for single-core
processors tend to load on one of the dies many of the tasks
that depend on each other. This makes recovery of tasks more
time-consuming.
In this paper, we propose a task scheduling algorithm
that makes recovery from a saved state faster on multicore
processor systems. The proposed method is based on the
scheduling algorithm proposed by Sinnen et al.[1], which takes
into account network contention and is capable of generating
schedules that are highly reproducible on real computing
systems. The proposed algorithm is designed as a parallelized
algorithm that can achieve O(n) times of speed-up when n
processors are available.
We evaluated our method with simulations and experiments
using 4 quad-core PCs, and the execution time including
recovery time can be reduced up to 50% in case of failure,
while the overhead in case of no failure was a few percent on
computation-heavy tasks.
The remainder of this paper is organized as follows. Section
II introduces some related works. Section III first explains the
key ideas and assumptions on the proposed method, and then
presents the proposed methods. Section IV shows the results
of the evaluation of execution times in case of failure and no
failure. Section V presents our conclusions.
II. RELATED WORKS
There are many kinds of task scheduling. In this paper, we
assume that task scheduling is assigning a processor to each
task, where the dependence of the tasks is represented by a
directed acyclic graph(DAG). The problem to find the optimal
schedule is NP-hard[4], [9], and there are many heuristic
algorithms for the problems[2], [3], [4], [8], [12].
List scheduling is a classical task scheduling method that
assigns the processor that can finish each task to the task in
order of a given priority of the tasks[13]. The priority can be
given by performing topological sorting on the dependence
graph of the tasks, for example.
Wolf et al. proposed a task scheduling method in which
multiple task graph templates are prepared beforehand, and
the assigned processors are determined according to which
templates fits the input task[5]. This method is suitable for
highly loaded systems, but does not take account of network
contention or processor failure.
© 2012 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including
reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any
copyrighted component of this work in other works. DOI:10.1109/CCGrid.2012.23

In many task scheduling methods, there is no consider-
ation of network contention. In these methods, a network
is assumed to be ideal communication paths where there is
no bandwidth limitation or latency, and sometimes schedules
generated by these methods cannot be well reproduced on real
computer systems. Sinnen et al. made a network contention
model and proposed a scheduling method based on it [1]. In
their experiments, improvement of accuracy and efficiency of
system utilization are shown by generating schedules while
considering network contentions. In their method, failure of
computing devices is not considered.
Checkpointing is a method for recovering task execution
in case of failure of computing nodes. Gu et al. proposed a
method for recovery from failure based on checkpointing in
stream processing systems[6], [7]. In this method, a state of
a node is saved when execution and transfer of the resulting
data of each task node to the subsequent processing node is
completed. In case of processor failure, it finds the closest
ancestor node which is not affected by the failure and recovers
the processing results from the saved state in that node.
Park et al. proposed a scheduling algorithm called team
scheduling that assigns a stream program to a multicore
architecture[11]. This scheduling improves control over buffer
space and achieves lower latency and deadlock-free feedback
loops while maintaining similar synchronization overhead to
the existing methods.
As far as we surveyed, there is no scheduling method
that takes account of multicore processor failure and network
contention. In this paper, we formulate the problem of task
scheduling for multicore processor systems that reduces re-
covery time from processor failure taking account of network
contention, and propose a heuristic method for the problem.
III. PROPOSED METHOD
In this section, we first describe the key idea of our proposal,
and define the terms used in this paper. Then, we describe
assumptions made in this paper and formulate the problem.
After that, we explain the proposed method.
A. Key Idea
A task graph is a DAG that represents the dependence
between tasks in a group of tasks. Fig.1(a) shows an example
of a task graph consisting of 3 task nodes 1, 2 and 3. It shows
that tasks 1 and 2 can be executed in parallel, while task 3
can only be started after outputs from tasks 1 and 2 have
been fully delivered to the processor that executes task 3. A
(a) (b)
Fig. 1. (a) Example task graph (b) Example processor graph
processor graph is a graph that represents the topology of a
network between processors. Fig.1(b) is an example processor
graph showing that 4 processors are connected by 3 network
switches. In this paper, we model a dual core processor as a
combination of two processors and a switch that connects these
processor cores. Thus, this graph can be viewed as a network
connecting 2 dual core processors with one network switch.
Task scheduling is generating an assignment of a processor
graph node to each of the task graph nodes.
Fig.2(a) shows a schedule generated by a method that does
not take account of processor failure, from the task graphs
shown in Fig.1(a) and Fig.1(b). In this schedule, task nodes
1 and 2 are assigned to processors A and B that are on
a same die, and task node 3 is also assigned to processor
A, since the communication between processors on a die is
significantly faster than that between processors on different
dies. The bottom figure in Fig.2(a) shows the schedule in this
case, and it represents that task nodes 1 and 2 are executed by
processor nodes A and B in parallel. After that, task node 3 is
executed on processor node A. As we described in Section II,
we can prepare for processor failure by saving the state to the
computing node after finishing execution of each task node.
However, in systems that utilize multicore processors, one
multicore processor failure results in simultaneous multiple
processor failure, and this makes the problem more complex.
In the example shown in Fig.2, if the die on which processor
nodes A and B are implemented fails when task node 3 is
being executed, all of the results of execution of task nodes
1, 2 and 3 are lost. In order to recover task nodes 1 to 3 in
this case, we need to execute task nodes 1 to 3 again, which
more than doubles the execution time, including recovery, that
is required in the case of no failure.
Fig.3 shows the schedule generated by the proposed method.
If a processor failure occurs in the same timing as described
above, we only lose the execution result of task node 3. Since
the results by execution of task nodes 1 and 2 are stored in the
computing node that houses processor nodes A and B, we can
(a) (b)
Fig. 2. (a) Schedule by existing method (b) Recovery with existing method

(a) (b)
Fig. 3. (a) Schedule by proposed method (b) Recovery with proposed method
use these results to re-execute task node 3 on processor node
A, and in this way we can make a quick recovery. On the other
hand, if no failure happens, we need extra communication time
between two dies. This imposes overhead on the proposed
method. We will explain how to minimize this overhead.
B. Definition
Here, we define terms used in this paper. The symbols used
are listed in Table I.
Task graph A task graph is a DAG in which each node
represents a task to be performed. The computation time to
execute task v, which is a node in the graph, is denoted
w(v). A directed arc in the graph is called a task link, and
a task link from node va to vb indicates that task va must be
completed before task vb begins. A task link e also represents
communication between two nodes, and if the data transfer is
performed using an inter-die communication link, it requires
c(e) length of time in order to finish the data transfer. The set
of all task nodes and the set of all task links are denoted V and
E, respectively. The task graph is denoted G = (V, E, w, c). In
this paper, the network model is based on the model proposed
by Sinnen et al.[1], and we assume that the following two
conditions are satisfied: (i) one processor can execute only one
task at a time, (ii) A task node cannot be executed until all
TABLE I
SYMBOLS USED IN THIS PAPER
Symbol Meaning
G Task graph
V Set of all task nodes
E Set of all task links
H Processor graph
P Set of all processor nodes
R Set of all processor links
eij Task link from i-th task node to j-th task node
proc(n) Processor assigned to task node n ∈ V
Proc(N) Set of processors assigned to task nodes in set N ⊆ V
pred(ni) Set of all parent nodes of ni
execution of parent nodes and all corresponding data transfers
are finished. Hereafter, the i-th task node is denoted ni. The
set of all parent nodes of a node n is denoted pred(n).
Processor graph A processor graph is a graph that represents
the network topology between processors. A node with only
one link is called a processor node. A processor node corre-
sponds to one processor core. A node with two or more links is
called a switch. A switch is not capable of executing a task but
only relays communication. An edge is called a processor link,
and it represents a communication link between processors and
switches. One multicore processor is represented by multiple
processor nodes, a switch and processor links connecting them.
The set of all processor nodes and the set of all processor
links are denoted P and R, respectively. The processor graph
is denoted H = (P, R).
C. Assumptions
In this paper, we make the following assumptions.
Each processor node saves the state to the computing node
housing that processor node when each task is finished. A
state contains the data transferred to the processor nodes that
will execute the child nodes of the completed task node. If a
processor failure occurs, a saved state which is not affected
by the failure is found in the ancestor task nodes. For the lost
results, some tasks are executed again using the saved state.
We assume that only output data from each task node is saved
as a state, which is much smaller than the complete memory
image, as in [6]. Thus, saving a state can be performed by
just replicating the output data on the main memory, and we
assume that saving a state does not consume time or resources.
We assume that only a single fail-stop failure of a multicore
processor can occur. When there is a failure of a multicore
processor, all processor cores on the die just stop execution
simultaneously and never perform a faulty operation. All states
saved in the computing node that houses the failed multicore
processor are lost in the failure. A failure can be detected by
interruption of heartbeat signals. Heartbeat signals are reliably
delivered with the highest priority without delay, and a failure
is detected within 1 second.
We assume that the failed computing node automatically
restart, and after predetermined time (about 1 minute), all
of the failed multicore processor comes back online. When
a multicore processor is rebooting, it cannot execute a task or
perform communication.
When data transfer is performed over network links be-
tween two processor nodes, due to bandwidth limitation these
network links cannot perform other data transfers. This is
called network contention. In this paper, we use the network
contention model proposed by Sinnen et al.[1], and we assume
the following conditions are satisfied: if data are transferred
through a series of processor links, downstream links cannot
start data transfer before upstream links; communication inside
a same die finishes instantly; all processors on a same die
share a network interface that can be used to communicate
with devices outside the die; all communication links outside
dies have the same bandwidth.

D. Problem Definition
Here, the length of schedule s is denoted sl(s). Suppose
that when a schedule s is executed, a fault happens at task
node v, and we make the schedule for the remaining tasks by
Sinnen’s scheduling algorithm. The entire schedule, including
the schedule before and after recovery, is denoted rs(s, v).
As inputs to the problem, a task graph G = (V, E, w, c) and
a processor graph H = (P, R) and which processor belongs
to each die are given. Our scheduling problem is to find the
schedule s such that:
minimize max
v∈V
sl(rs(s, v)).
E. Proposed Algorithm
We first explain the overview of the proposed method, and
then explain the details using pseudocodes.
The critical path in a task graph is the path from the first
task to the last task which has the longest execution time, and
this influences greatly the determining of the schedule length.
If all the tasks in the critical path are assigned to processors on
a die, and a fault happens at the last task in the critical path,
the entire execution time of the schedule is doubled, and this
is the worst case. We can distribute the tasks in the critical
path over multiple dies in order to improve recovery time,
but this could increase the overhead in case of no failures,
due to extra communication between dies. On the other hand,
the last task in the critical path has the largest influence on
the recovery time. We therefore distribute the last nmove
tasks in the critical path over multiple dies and try to find
the best nmove. Since the proposed algorithm only tries to
distribute the tasks in the critical path, it can quickly generate
the schedule.
Algorithm 1 List scheduling
INPUT: Task graph G = (V, E, w, c) and processor graph
H = (P, R).
1: Sort nodes n ∈ V into list L, according to priority scheme
and precedence constraints.
2: for each n ∈ L do do
3: Find processor p ∈ P that allows earliest finish time of
n.
4: Schedule n on p.
5: end for
The classical list scheduling algorithm is shown in Al-
gorithm 1. In the list scheduling, each task is assigned to
the processor that allows the earliest finish time of the task,
in descending order of bl, that is the length of remaining
schedule. In the classical list scheduling, network contention
is not considered.
The algorithm proposed by Sinnen, et al. is shown in
Algorithm 2, which is an enhancement of the list scheduling
algorithm by adding network scheduling. In this algorithm,
Algorithm 3 is called when findProcessor function is called.
Below, we give some explanation for the pseudocode.
Algorithm 2 Scheduling considering network contention
H = (P, R).
1: Sort nodes nj ∈ V into list L in descending order of bl,
according to precedence constraints.
2: for each n ∈ L do do
3: Find processor p = findProcessor(P, nj).
4: for each ni ∈ pred(nj) in a definite order do
5: if proc(ni) = p then then
6: determine route R = [L1, L2, ..., Ll] from
proc(ni) to p.
7: schedule eij on R.
8: end if
9: end for
10: schedule nj on p.
11: end for
12: return the schedule.
Algorithm 3 Processor selection considering network con-
tention
INPUT: Task node n and set of processors P.
1: return p ∈ P that allows earliest finish time of n, taking
account of network bandwidth usage.
Line 2 to 11: Each task node nj ∈ L is assigned a processor
in order of the position in L.
Line 3: The processor assigned to nj is determined by find-
Processor function. Reserved bandwidth in line 7 is referred
here.
Line 4 to 9: Bandwidth of eij is reserved for the network route
between the processor assigned to ni (which is the parent node
of nj) to the processor assigned to nj.
The pseudocode for findProcessor function in Algorithm 2
is shown in Algorithm 3, and this function corresponds to
line 3 in Algorithm 1. Unlike the classical list scheduling, a
processor is selected according to both the computation time
and the communication time considering bandwidth reserved
at line 7 in Algorithm 2.
The proposed algorithm is shown in Algorithm 4. When
Algorithm 2 is called at line 3, Algorithm 5 is used as a
findProcessor function. Algorithm 5 selects the processor that
can finish the given task first, except the processor to which
the parent task is assigned. Since the proposed algorithm is
based on Algorithm 2, it generates schedules taking account
of network contention. Since the loops in Algorithm 4 are
independent, these loops can be executed in parallel. We gen-
erate the schedule for recovery after a fault using Algorithm
2 and Algorithm 3.
The pseudocode for the proposed method is shown in Algo-
rithm 4. Below, we give some explanation for the pseudocode.
Line 2 to 13: nmove is the number of task nodes moved to
another die. For each of all task nodes, we generate a schedule
where a failure occurs during execution and a recovery is
performed. We find the longest schedule from them, and find

Algorithm 4 Proposed method
H = (P, R).
1: F = ∅.
2: for (in parallel) nmove = 1 ; nmove ≤ number of tasks
in critical path do
3: s1 = Schedule generated by Algorithm 2 and Algorithm
5. (refer to the body text)
4: S = ∅.
5: for (in parallel) failtask = 1 ; failtask ≤ number
of all tasks do
6: Find set T of all tasks that will be executed after
recovery if task failtask fails.
7: s2 = schedule generated by Algorithm 2 and Algo-
rithm 3 with T as the given task nodes and processors
are available only after finishing current tasks or
rebooting.
8: s = concatenated schedule of s1 and s2
9: S = S {s}.
10: end for
11: f = the longest schedule in S.
12: F = F {f}.
13: end for
14: Find the shortest schedule g in F.
15: Let s3 be the original schedule without failure correspond-
ing to g.
16: return s3.
nmove that makes the longest schedule shortest.
Line 3: Here, Algorithm 2 is called, and when findProcessor
is called, Algorithm 5 is executed instead of Algorithm 3.
Algorithm 5 selects a processor so that task nodes are not
concentrated on a single die. We describe the details later.
Line 5 to 10: For task node failtask, a schedule is generated
assuming that a failure occurs at failtask and recovery is
performed, by concatenating the schedule before failure and
the schedule for the recovery.
Line 7: Using Algorithm 2, recovery schedule is generated.
Here, Algorithm 3 is executed when findProcessor function
is called in Algorithm 2.
Since the iteration from line 2 to line 13 and line 5 to line 10
are independent from each other, these loops can be executed
in parallel.
The pseudocode for the processor selection in the proposed
method is shown in Algorithm 4. Below, we give some
explanation for the pseudocode.
Line 2: Checks if task node n is in the last nmove tasks in
the critical path.
Line 3 to 7: Find set of processors Proc(pred(n)) that
are assigned to the set of task nodes pred(n). pred(n) is
the set of parent nodes of n. Then, select the processor
that allows earliest finish of n from the processors excluding
Proc(pred(n)).
Line 4: If there exist dies excluding Proc(pred(n)), then
Algorithm 5 Processor selection for proposed method
INPUT: Task node n and set of processor nodes P. nmove
is defined in Algorithm 4.
1: if pred(n) = ∅ then
2: if n is in the first nmove tasks in the critical path then
3: if there is a processor which is on different die than
the proc(pred(n)) then
4: return the processor that can finish the task except
processors on the same die as Proc(pred(n)),
considering network contention.
5: else
6: return p ∈ P that allows earliest finish time of
n, considering unused bandwidth between p and
Proc(pred(n)).
7: end if
8: else
9: return p ∈ P that allows earliest finish time of
n, considering unused bandwidth between p and
Proc(pred(n)).
10: end if
11: end if
select the processor that allows earliest finish of n from them.
Line 6: If there is no die except Proc(pred(n)), then select
the processor that allows earliest finish of n from all proces-
sors.
Line 9: If task node n is not in the last nmove tasks in the
critical path, then select the processor that allows earliest finish
of n from all processors.
Since the proposed method is based on the algorithm
proposed by Sinnen et al, it allows scheduling taking account
of network contention.
IV. EVALUATION
In order to evaluate the overhead and to seek improvement
of recovery time in case of failure, we compared our method
with two other methods using a real computing system con-
sisting of 4 PCs. We compared the following three methods.
Contention The scheduling algorithm proposed by Sinnen, et
al.[1], that takes account of network contention. After a failure,
the task graph that contains the remaining task nodes that need
to be executed including the task nodes whose results are lost
is generated. Then a schedule for this task graph is generated
using Sinnen’s method.
Proposed The proposed method. The schedule after failure is
generated in the same way as Contention method.
Interleaved This scheduling algorithm tries to spread tasks to
all dies as much as possible. This algorithm is also based on
Sinnen’s algorithm, but it assigns tasks to all dies in turn. The
problem of existing methods to deal with a failure is that they
tend to assign dependent tasks to one die, and this algorithm is
intended to improve this situation. However, this does not take
account of the efficiency in case of no failure, where a large
overhead is expected. The schedule after failure is generated
in the same way as with the Contention method.

We used Robot Control, Sparse Matrix Solver and various
random graphs from the Standard Task Graph Set[10], [14] as
the input task graph to the three methods.
A. Environment
We used a computing system consisting of 4 PCs connected
with Gigabit Ethernet. Each PC has an Intel Core i7 920
(2.67GHz) as the CPU, Intel Gigabit CT Desktop Adaptor
EXPI9301CT(PCI Express x1) as the network interface card
and 6.0G bytes of memory. We measured execution time
with an originally developed program running on Java(TM)
SE Runtime Environment (build 1.6.0_21-b07, 64bit) and
Windows 7 Enterprise (64bit). We used the standard TCP
sockets to transfer data between these PCs.
We used the processor graph that represents the above
computing system. We assumed that each PC has 4 processors,
and we did not count the hardware threads provided by hyper-
threading. We used 450Mbps as the bandwidth of processor
links outside the dies, which is the measured value using the
above system.
We measured the time for Ubuntu 10.04 operating system
running on VMware to reboot after pushing the reset but-
ton, and it took 25 seconds on average. Thus, we used 25
seconds as the time required for rebooting after failure. In
the experiments, we reproduced the failure by suppressing the
corresponding processors to perform computation or commu-
nication for 25 seconds, instead of performing a hard reset
on the computers. We set the experiments so that each failure
would be detected in 1 second.
In order to determine the worst timing of failure, we ﬁrst
made the schedule by Sinnen’s method and found the task
that takes the largest execution time including recovery when
a failure occurs during execution of the task. We made the
system fail when this task is being executed.
B. Sparse Matrix Solver
Fig.6(a) shows the task graph for the Sparse Matrix Solver
that has 98 tasks and 177 links. This graph represents a sparse
matrix solver of an electronic circuit simulation generated by
the OSCAR FORTRAN compiler. This graph has a relatively
high level of parallelism.
Fig.4 and Fig.5 show the results of experiments in case of
failure and no failure. The graphs show the measured results
on the system with the three methods and the simulated results
for the proposed method. We can see that there are some
difference between the simulated results and the results on
the real system. This is mainly because of the instability of
network bandwidth between PCs.
Fig.4 shows the execution time in case of failure. The
Proposed method has a shorter execution time than the other
two methods regardless of the number of dies (or PCs). In
the case of 4 dies, in fact, the Proposed and the Interleaved
methods, achieve signiﬁcantly better execution time than the
Contention method.
Fig.5 shows the execution time in case of no failure.
The Proposed method has better execution time than the
(a) (b)
Fig. 6. (a) Sparse Matrix Solver (b) Task graph for Robot Control
Interleaved method, and the maximum overhead against the
Contention method was 12% at maximum.
C. Robot Control
Fig. 6(b) shows the task graph for the Robot Control
that has 90 tasks and 135 edges. This graph represents a
Newton-Euler dynamic control calculation for the 6-degrees-
of-freedom Stanford manipulator, which has a lower level
of parallelism compared to the Sparse Matrix Solver. The
experimental results for this graph are shown in Fig. 7 and
Fig.8. In this experiment, we used only 2 cores in each PC,
since the task graph does not have enough size for using all
4 cores in the PCs.
Fig.7 shows the results in case of failure. Proposed method
has the best execution time regardless of the number of dies.
When the number of dies is 4, in fact, the Proposed method
achieves more than 50% of speed-up against the Contention
method.
Fig.8 shows the results in case of no failure. The maximum
overhead for the Proposed method against the Contention
method was 10% at the maximum. The Interleaved method
has 30% overhead, which is much larger than the Proposed
method. The Proposed method always had smaller overhead
than the Interleaved method in this experiment.
The Contention method generates schedules that execute
fast when there is no failure. However, many execution results
of task nodes are lost when a failure occurs, and that increases
the time for recovery. Especially when there are many dies,
the Contention method requires larger recovery time.
In case of no failure, the Proposed method has larger
overhead than the Contention method. This is due to the extra
communication between dies. However, the overhead is much
less than that of the Interleaved method, which, as much as
possible, spreads tasks over dies.
D. Random Graphs
We used various random graphs with 50 task nodes. The
number of links and the level of parallelism differ for each
graph. We also varied the CCR of the graphs, where CCR is

Fig. 4. Sparse matrix solver w/failure Fig. 5. Sparse matrix solver, no failure
Fig. 7. Robot Control w/failure Fig. 8. Robot Control, no failure
Fig. 9. Random Task Graph w/failure Fig. 10. Random Task Graph, no failure

the total communication time divided by the total computation
time, and a large CCR means a communication-heavy task. We
measured the execution times when CCR is 0.1, 1 and 10. In
this experiment, we fixed the number of dies to 4 and each
die has 4 cores.
Fig.9 shows the results in case of failure. The Proposed
method has a better execution time than the Contention method
regardless of CCR. When CCR is 0.1, the Interleaved method
is slightly better than the Proposed method. This is because in
computation-heavy tasks, spreading tasks over processor dies
is sometimes very effective. However, when CCR is 1 or larger,
the Proposed method was the best.
Fig.10 shows the results in case of no failure. This graph
shows that as CCR increases, the overhead for the Proposed
method and Interleved method increases. This is because the
overhead basically comes from the communication time, and
more communication is needed in communication-heavy tasks
to spread tasks over processor dies. This is especially notable
in Interleved method. On the other hand, even when CCR is 10,
the overhead for Proposed method was 35%, and it succeeded
to improve the result by 6% in case of failure.
Fig. 11. Schedule generation time
E. Schedule Generation Time
Fig.11 shows the time taken to generate schedules for each
task graph when the scheduling algorithm is executed with a
single thread and with 8 threads. We used a PC with Intel Core
i7 920 processor with hyper threading to measure these results.
The proposed algorithm achieved up to 4 times of speed-up.
V. CONCLUSION
In this paper, we formalized a task scheduling problem for
multicore processor systems that reduces recovery time from
checkpoints when a single fail-stop failure occurs. We also
proposed a scheduling algorithm based on the existing method
that takes account of network contention. The basic idea of the
proposed scheduling algorithm is to suppress the scheduler in
order to assign dependent tasks to processors on a same die.
The proposed scheduling algorithm is designed as a parallel
algorithm that can achieve O(n) speed-up if n processors are
available.
We evaluated our proposed method on a computing system
consisting of 4 quad-core PCs connected via Gigabit Ethernet.
We confirmed that the proposed method can improve recovery
time by up to 50 percent while the overhead in case of no
failure was a few percent, in typical scenarios with a 1 or less
communication-to-computation ratio.
As future work, we would like to address the case where
the processing time of each task node cannot be known
beforehand. We are also planning to evaluate our algorithm
on larger computing systems.
REFERENCES
[1] Sinnen, O. and Sousa, L.A. : “Communication Contention in Task
Scheduling,” IEEE Trans. on Parallel and Distributed Systems, 16, 6,
pp. 503-515, 2005.
[2] Sinnen, O., To, A., and Kaur, M. : “Contention-Aware Scheduling with
Task Duplication,” Proc. of JSSPP’2009, pp. 157-168, 2009.
[3] Sinnen, O. Sousa, L.A. Sandnes, F.E. : “Toward a realistic task scheduling
model,” IEEE Trans. on Parallel and Distributed Systems, 17, 3, pp. 263-
275, 2006.
[4] Amato, N.M. and An, P. : “Task Scheduling and Parallel Mesh-Sweeps
in Transport Computations,” Technical Report, TR00-009, Department of
Computer Science and Engineering, Texas A&M University, 2000.
[5] Wolf, J., Bansal, N., Hildrum, K., Parekh, S., Rajan, D., Wagle, R., Wu,
K-L., and Fleischer, L. : “SODA: An Optimizing Scheduler for Large-
Scale Stream-Based Distributed Computer Systems,” Proc. of the 9th
ACM/IFIP/USENIX International Conference on Middleware, pp.306-
325, 2008.
[6] Gu, Y., Zhang, Z., Ye, F., Yang, H., Kim, M., Lei, H., and Liu, Z. : “An
empirical study of high availability in stream processing systems,” Proc.
of the 10th ACM/IFIP/USENIX International Conference on Middleware,
pp.23:1-23:9, 2009.
[7] Zhang, Z., Gu, Y., Ye F., Yang H., Kim M., Lei H., and Liu Z. : “A Hybrid
Approach to High Availability in Stream Processing Systems,” Proc.
of 2010 International Conference on Distributed Computing Systems,
pp.138-148, 2010.
[8] Guo, L., Sun, H., and Luo, Z. : “A Data Distribution Aware Task Schedul-
ing Strategy for MapReduce System,” Proc. of the 1st International
Conference on Cloud Computing, pp.694-699, 2009.
[9] Tang, X., Li, K., and Padua, D.A. : “Communication contention in
APN list scheduling algorithm,” Science in China Series F: Information
Sciences, pp.59-69, 2009.
[10] Tobita, T. and Kasahara, H. : “A standard task graph set for fair eval-
uation of multiprocessor scheduling algorithms,” Journal of Scheduling,
5, 5, pp.379-394, 2002.
[11] Park, J. and Dally, W.J. : “Buffer-space Efficient and Deadlock-free
Scheduling of Stream Applications on Multi-core Architectures,” Proc. of
the 22nd ACM symposium on Parallelism in algorithms and architectures,
pp.1-10, 2010.
[12] Macey, B.S. and Zomaya, A.Y. : “A performance evaluation of CP list
scheduling heuristics for communication intensive task graphs,” Proc. of
the First Merged Int’l Parallel Processing Symposium and Symposium
on Parallel and Distributed Processing 1998, pp.538-541, 1998.
[13] Sinnen, O. : “Task Scheduling for Parallel Systems,” Wiley Series on
Parallel and Distributed Computing, 2007.
[14] "Standard Task Graph Set" Home Page
http://www.kasahara.elec.waseda.ac.jp/schedule/

(Paper) Task scheduling algorithm for multicore processor system for minimizing recovery time in case of single node fault

More Related Content

What's hot (18)

Similar to (Paper) Task scheduling algorithm for multicore processor system for minimizing recovery time in case of single node fault (20)

More from Naoki Shibata (20)

Recently uploaded (20)

(Paper) Task scheduling algorithm for multicore processor system for minimizing recovery time in case of single node fault