Job Scheduling Strategies for Networks of
Workstations
1 1 2 3
B B Zhou R P Brent D Walsh and K Suzaki
1
Computer Sciences Lab oratory Australian National University
Canb erra ACT Australia
2
CAP Research Program Australian National University
Canb erra ACT Australia
3
Electrotechnical Lab oratory Umezono
sukuba Ibaraki Japan
Abstract In this pap er we rst intro duce the concepts of utilisation ra
tio and e ective sp eedup and their relations to the system p erformance
We then describ e a two level scheduling scheme which can b e used to
achieve go o d p erformance for parallel jobs and go o d resp onse for inter
active sequential jobs and also to balance b oth parallel and sequential
workloads The two level scheduling can b e implemented by intro ducing
on each pro cessor a registration o ce We also intro duce a lo ose gang
scheduling scheme This scheme is scalable and has many advantages
over existing explicit and implicit coscheduling schemes for scheduling
parallel jobs under a time sharing environment
Intro duction
The trend of parallel computer developments is toward networks of worksta
tions or scalable paral lel systems In this typ e of system each pro cessor
having a high sp eed pro cessing element a large memory space and full function
ality of a standard op erating system can op erate as a stand alone workstation
for sequential computing Interconnected by high bandwidth and low latency
networks the pro cessors can also b e used for parallel computing To establish
a truly general purp ose and user friendly system one of the main problems is
to provide users with a single system image By adopting the technique of dis
tributed shared memory for example we can provide a single addressing
space for the whole system so that communication for transferring data b etween
pro cessors is completely transparent to the client programs In this pap er we
discuss another very imp ortant issue relating to the provision of single system
image that is e ective job scheduling strategies for b oth sequential and paral lel
processing on networks of workstations
Many job scheduling schemes have b een intro duced in the literature and
some of them implemented on commercial parallel systems These scheduling
schemes for parallel systems can b e classi ed into either space sharing or time
sharing or a combination of b oth With space sharing a system is partitioned
into subsystems each containing a subset of pro cessors There are b oundary
lines laid b etween subsystems and so only pro cessors of the same subsystem
can b e co ordinated to solve problems assigned to that subsystem During the
computation each subsystem is allo cated only for a single job at a time
The space partition can b e either static or adaptive With static partitioning
the system con guration is determined b efore the system starts op erating The
whole system has to b e stopp ed when the system needs to b e recon gured
With adaptive partitioning pro cessors in the system are not divided b efore the
computation When a new job arrives a job manager in the system rst lo cates
idle pro cessors and then allo cates certain numb er of those idle pro cessors to
that job according to some pro cessor allo cation p olicies e g those describ ed
in Therefore the b oundary lines are drawn during the
computation and will disapp ear after the job is terminated Normally the static
partitioning is used for very large systems while the adaptive partitioning is
adopted in systems or subsystems of small to medium size One disadvantage of
space partitioning is that short jobs can easily b e blo cked by long ones for a long
time b efore b eing executed However in practice short jobs usually demand a
short turnaround time To alleviate this problem jobs can b e group ed into classes
and a sp ecial treatment will b e given to the class of short jobs However it
can only partially solve the problem Thus time sharing needs to b e considered
Many scheduling schemes for time sharing of a parallel system have b een
prop osed in the literature They may b e classi ed into two basic typ es The
rst one is local scheduling With lo cal scheduling there is only a single queue
on each pro cessor Except for higher or lower priorities b eing given pro cesses
asso ciated with parallel jobs are not distinguished from those asso ciated with
sequential jobs The metho d simply relies on existing lo cal schedulers on each
pro cessor to schedule parallel jobs Thus there is no guarantee that the pro cesses
b elonging to the same parallel job can b e executed at the same time across the
pro cessors When many parallel programs are simultaneously running on a sys
tem pro cesses b elonging to di erent jobs will comp ete for resources with each
other and then some pro cesses have to b e blo cked when communicating or syn
chronising with non scheduled pro cesses on other pro cessors This e ect can lead
to a great degradation in overall system p erformance One metho d
to alleviate this problem is to use two phase blo cking which is also called
implicit coscheduling in In this metho d a pro cess waiting for communication
spins for some time in the hop e that the pro cess to b e communicated with on
the other pro cessor is also scheduled and then blo cks if a resp onse has not b een
received The rep orted exp erimental results show that for parallel workloads this
scheduling scheme p erforms b etter than the simple lo cal scheduling However
the problem is that the scheduling p olicy is based on communication require
ments Then it tends to give sp ecial treatment to jobs with a high frequency of
communication demands The p olicy is also indep endent of service times The
p erformance of parallel computation is thus unpredictable
The second typ e of scheduling schemes for time sharing is coscheduling
or gang scheduling which may b e a b etter scheme in adopting short job rst
p olicy Using this metho d a numb er of parallel programs is allowed to enter a
service queue as long as the system has enough memory space The pro cesses
of the same job will run simultaneously across the pro cessors for only certain
amount of time which is called scheduling slot When a scheduling slot is ended
the pro cessors will context switch at the same time to give the service to pro
cesses of another job All programs in the service queue take turns to receive
the service in a co ordinated manner across the pro cessors Thus programs never
interfere with each other and short jobs are likely to b e completed more quickly
There are also certain drawbacks asso ciated with coscheduling A signi cant one
is that it is designed only for parallel workloads For networks of workstations we
need an e ective scheduling strategy for b oth sequential and parallel pro cessing
The simple coscheduling technique is not a suitable solution
The future networks of workstations should provide a programming free en
vironment to general users By providing a variety of high p erformance com
puting libraries for a wide range of applications plus user friendly interfaces for
the access to those libraries parallel computing will no longer b e considered
just as client s sp ecial requests but b ecome a natural and common phenomenon
in the system Along with many other critical issues therefore highly e ective
job management strategies are required for the system to meet various client s
requirements and to achieve high e ciency of resource utilisation Because of
the lack of e cient job scheduling strategies most networks of workstations are
currently used exclusively either as an MPP for pro cessing parallel batch jobs
or as a group of separate pro cessors for interactive sequential jobs The p oten
tial p ower of this typ e of system are not exploited e ectively and the system
resources are not utilised e ciently under these circumstances
In this pap er we discuss some new ideas for e ectively scheduling b oth se
quential and parallel workloads on networks of workstations To achieve a de
sired p erformance for a parallel job on a network of workstations with a variety
of comp etitive background workloads it is essential to provide a sustained ra
tio of CPU utilisation to the asso ciated pro cesses on each pro cessor to allo cate
more pro cessors to the job if the assigned utilisation ratio is small and then to
co ordinate the execution across the pro cessors We rst intro duce the concepts
of utilisation ratio and e ective speedup and their relations to the system p er
formance in Section In this section we also argue that b ecause the resources
in a system are limited one cannot guarantee every parallel job to have a sus
tained CPU utilisation ratio in a time sharing environment One way to solve
the problem is that we give short jobs sustained utilisation ratios to ensure a
short turnaround time while to each large job we allo cate a large numb er of pro
cessors and assign a utilisation ratio which can vary in a large range according
to the current system workload so that small jobs will not b e blo cked and the
resource utilisation can b e kept high we then present in Section a two level
scheduling scheme which can b e used to achieve go o d p erformance for paral
lel jobs and go o d resp onse for interactive sequential jobs and also to balance
b oth parallel and sequential workloads The two level scheduling can b e imple
mented by intro ducing on each pro cessor a registration o ce which is describ ed
in Section We discuss a scalable coscheduling scheme loose gang schedul
ing in Section This scheme requires b oth global and lo cal job managers It
is scalable b ecause the coscheduling is mainly controlled by lo cal job managers
on each pro cessor so that frequent signal broadcasting for simultaneous context
switch across the pro cessors is avoided Using a global job manager we b elieve
that the system can work more e ciently than those using only lo cal schedulers
With a lo cal job manager on each pro cessor the system will b ecome more exible
and more e ective in handling more complicated situations than those adopting
only the conventional gang scheduling p olicy Finally the conclusions are given
in Section
Utilisation Ratio and E ective Sp eedup
Assuming that the overall computational time for a parallel job on p dedicated
pro cessors is T p the conventional ly de ned speedup is then obtained as
d
T
d
S p
d
T p
d
This sp eedup can only b e achieved by using dedicated pro cessors It may b e
imp ossible to achieve on a network of workstations b ecause there a parallel job
usually has to time share resources with other sequential parallel jobs If we
provide a sustained ratio of CPU utilisation for a job on each pro cessor and use
more processors however we can still achieve the desired p erformance in terms
of time
De ne utilisation ratio for as the ratio of CPU utilisation for
a given job on each pro cessor By a given the job on a pro cessor can on the
average obtain a service time T in each unit of time T In our scheduling
strategy each parallel job will b e assigned a utilisation ratio which is usually
determined based on the current system working conditions Di erent ratios can
also b e given on di erent pro cessors for naturally unbalanced parallel jobs to
achieve b etter system load balancing
Assume that the same utilisation ratio is assigned to a parallel job across
all the asso ciated pro cessors and that the job s pro cesses are gang scheduled
The turnaround time T p for that job can then b e calculated as
e
T p
d
T p
e
where T p is the computational time obtained on p dedicated pro cessors
d
De ning e ective sp eedup S p as the ratio of T and T p then
e d e
T T
d d
S p S p
e d
T p T p
e d
where S p is the conventional sp eedup obtained on p dedicated pro cessors
d
To achieve a desired p erformance we may set a p erformance target and
require
T T p
d e
or
S p
e
If the e ective sp eedup for a given job is lower than that target the p erformance
will b e considered unacceptable
From equations in and we can obtain
S p
d
Using the ab ove inequality we can easily determine how many pro cessors
should b e allo cated to a given job in order to achieve a desired p erformance
when a particular is given Assuming and for example S p
d
must b e greater than or equal to Allo cating pro cessors or more to that job
can then achieve a desired p erformance if S When the current system
d
workload is not heavy we may need to use less numb er of pro cessors to achieve
the same p erformance If there are several idle pro cessors we may set in
the ab ove example Then only pro cessors may b e required if S
d
In practice the exact sp eedup S p may not b e known except for those
d
programs in standard general purp ose parallel computing libraries Thus the
values can only b e approximate in those cases However go o d approximations
can often b e obtained For example the results of the Linpack Benchmark
can b e used as a go o d approximation for problems of matrix computation
The utilisation ratios of the existing jobs may b e decreased whenever a new
job enters the system to time share the resources The problem is how to ensure
a sustained ratio of CPU utilisation for each job so that the p erformance can b e
predictable in a time sharing environment Since the resources in a system are
really limited the answer to this question is simply that we cannot guarantee
every job to have a sustained ratio when the system workload is heavy
One way to solve the ab ove problem is to adopt the following scheme First we
set a limit to the length of each scheduling round T or a limit to the numb er of
jobs in the system A common misunderstanding ab out time sharing for parallel
jobs is that go o d p erformance will b e obtained as long as parallel jobs can enter
the system and start op eration quickly As we mentioned previously that the
resources in a system are limited however go o d p erformance just cannot b e
guaranteed if the length of scheduling round is unb ounded Consider a simple
example when several large jobs are time sharing the resources in a round robin
manner In this case the conventional gang scheduling simply fail to pro duce
go o d p erformance in terms of turnaround time
Because of the limit to the length of each scheduling round short jobs still
can b e blo cked for a long time We then adopt a scheduling p olicy that is small
jobs should have sustained utilisation ratios to ensure a short turnaround time
while each large job should b e assigned a large numb er of pro cessors but given a
utilisation ratio which can vary in a large range according to the current system
workload In this way we think that small jobs will not b e blo cked the resource
utilisation can b e kept high and reasonably go o d p erformance for large jobs may also b e obtained
Based on the ab ove ideas a multi class time space sharing system is designed
A detailed description of this system is b eyond the scop e of this pap er Interested
readers may refer for more details
Two Level Scheduling
It can b e seen from the previous section that our scheduling strategy is based
on the utilisation ratios assigned to parallel jobs In this section we intro duce a
two level scheduling scheme for balancing the workloads for b oth sequential and
parallel pro cessing
At the top level or global level the gang scheduling or a loose gang scheduling
scheme to b e discussed in the next section is adopted to co ordinate parallel
computing Each scheduling round T is divided into time slots An example
of the time distribution for di erent pro cesses on each pro cessor is shown in
(i)
Fig In the gure time slot t is allo cated only to sequential processes
s
(i)
is assigned to a single paral lel asso ciated with sequential jobs while slot t
p
process asso ciated with a parallel job A parallel pro cess may share its time
slots with sequential pro cesses through the scheduling at the b ottom level or
local level However no parallel pro cesses will share the same time slots This
is to avoid many di erent typ es of parallel jobs comp eting for resources at the
same time and then to guarantee that each parallel pro cess can obtain its prop er
share of resources The relation b etween a scheduling round and those time slots
satis es the following equation
n
X
(i)
t T T
s
p
i=1
P
(i)
m
where T t is the total time dedicated for sequential jobs in a
s
s
i=1
scheduling round and is distributed to gain go o d resp onse to interactive clients
(1) (2) (3) (4)
(1) (2) (3) (4)
t t t t
p p p p
t t t t
s s s s
T
Fig The time distribution in a scheduling round
The width of each time slot is determined by the corresp onding utilisation
(i) (i)
ratio or We can then calculate the width of each time slot as
p s
(i) (i)
T t
p p
and
T T
s s
P
(i)
m
where
s
s
i=1
There are many ways to distribute T For example each slot for a parallel
s
pro cess can b e followed by a small slot for sequential pro cesses and T is
s
uniformly distributed across the whole scheduling round Then
T
s
(i)
t
s
n
We can also distribute T proportional ly to the width of each time slot for
s
parallel pro cesses that is
(i)
p
(i)
T t
s
P
s
(j )
n
p
j =1
The calculation for proportional distribution is a bit more complicated than
that for uniform distribution However it is useful when a proper share p olicy
which will b e describ ed later in the section is applied at the lo cal level
Di erent lo cal p olicies can b e adopted to schedule pro cesses within each time
slot In those time slots dedicated for sequential pro cessing conventional lo cal
scheduling schemes of any standard op erating system will b e go o d enough In
(i)
the following we discuss how to schedule pro cesses in each time slot t in
p
which parallel pro cessing is involved
To ensure that a parallel pro cess can obtain its assinged share of CPU util
(i)
isation the whole slot t may b e dedicated just to the asso ciated parallel
p
pro cess In that case a very high priority will b e given and the pro cess simply
do es busy waiting or spins during communication synchronisation so that no
other pro cesses can disturb its execution within each asso ciated time slot One
problem asso ciated with this p olicy is that the p erformance of sequential jobs
esp ecially of those which demand go o d interactive resp onse may signi cantly b e
a ected Therefore its use will b e treated as sp ecial cases under the environment
of networks of workstations to achieve certain client s sp ecial requests
To prevent great p erformance degradation of sequential interactive jobs im
plicit coscheduling scheme can b e adopted However a p otential problem is that
the execution of a parallel pro cess may b e disturb ed by several sequential pro
cesses and then it is p ossible that certain parallel pro cesses may not receive their
prop er shares in their asso ciated time slots
The ab ove problem may b e alleviated by adopting a proper share p olicy In
this p olicy we do not consider individual shares allo cated for each sequential job
Except for sp ecial ones e g multimedia workloads which may b e treated in the
same way as parallel jobs to achieve constant rate services only a combined
(i)
share of sequential pro cesses t is considered Each distributed time slot for
s
(i) (i)
to is also integrated with its asso ciated time slot t sequential pro cesses t
p s
(i)
form a single time slot of width t that is
(i) (i) (i)
t t t
s p
In each integrated time slot implicit coscheduling is applied to supp ort b oth
parallel and sequential pro cessing When its allo cated share is not used up in
(i)
time t a parallel pro cess can still obtain services till the end of the integrated
p
(i)
(i) (i)
time slot t though t is longer than t When a parallel pro cess has
p
consumed its share b efore the end of an integrated time slot however it will b e
blo cked and the services in the remaining time slot then dedicated to sequential
pro cesses With this p olicy parallel pro cesses and sequential pro cesses as a whole
may b e guaranteed to obtain their prop er shares during the computation
Similar to the one describ ed in the p olicy may b e realised by applying
the proportional share technique which are originally used for real time appli
cations However our scheduling scheme is much simpler and easier to
implement b ecause only the prop er share of a single parallel pro cess is considered
against a combined share of sequential pro cesses in each time slot
Now the problem is how to distribute the total time T allo cated for pro
s
cessing sequential jobs The uniform distribution using the equation in is
(i)
easy to calculate However the resulting t may b e to o small to comp en
s
(i)
sate the lost share of parallel pro cesses which have large s Therefore the
p
prop ortional distribution using may b e a more prop er one
Normalising T that is setting T the equation in will b ecome
n
X
(i)
s
p
i=1
Using equations in and we obtain
(i)
p
(i) (i) (i)
T t t t
P
s p
(j )
n
p
j =1
(i)
The width of an integrated time slot t can directly b e obtained by using
(i)
the equation in and thus there is no need to explicitly calculate t s
s
Registration O ce
When a parallel pro cess has used up its time slot it will b e preempted at the
global level and another parallel pro cess b e dispatched After b eing dispatched
parallel pro cesses may time share resources with sequential pro cesses on each
pro cessor Just like sequential pro cesses parallel pro cesses will then b e either
in running state or ready and blocked states which is controlled by a lo cal
scheduler Because in our two level scheduling the execution of parallel pro cesses
are controlled at b oth global and lo cal levels sp ecial care has to b e taken to
avoid p otential scheduling con icts For example the global scheduler wants to
preempt a parallel pro cess which is currently not in running state To solve this
problem we intro duce a registration o ce on each pro cessor
The registration o ce is constructed by using a linked list as shown in Fig
When a parallel job is initiated each asso ciated pro cess will enter the lo cal se
quential queueing system the same way as sequential pro cesses on the corre
sp onding pro cessor Just like sequential pro cesses parallel pro cesses can b e ei
ther in running state or in ready state requesting for service or in blo cked state
manager servant
alg
timer
P
registration o ce
pro cess dispatched
H no de no de no de no de T
pro cs pro cs pro cs pro cs
IN OUT IN IN
Fig The organisation of a registration o ce
during communication synchronisation However every parallel pro cess has to
b e registered in the registration o ce that is on each pro cessor the linked list
will b e extended with a new no de which has a p ointer p ointing to the pro cess
just b eing initiated Similarly when a parallel job is terminated it has to check
out from the o ce that is the corresp onding no de on each pro cessor will b e
deleted from the linked list
As we discussed in the previous section certain parallel pro cesses may b e as
signed a very high priority so that they can o ccupy the whole time slots allo cated
to them In that case the execution of sequential workloads can b e seriously de
(i)
teriorated To alleviate this problem we may intro duce certain time slots t
s
which are dedicated to sequential jobs only This can b e done by intro ducing
dummy nodes in the linked list A dummy no de is the same typ e of no des in a
linked list except its p ointer p oints to NULL the constant zero instead of a real
parallel pro cess It seems that there is a dummy paral lel process asso ciated with
that no de When a service is given to that dummy parallel pro cess the whole
time slot will b e dedicated to sequential pro cesses
There is a servant working in the o ce When the servant comes to a place
or a no de in the linked list the pro cess asso ciated with that no de can receive
services or be dispatched When a pro cess is dispatched it will b e marked out
Other pro cesses which are not dispatched will b e marked in In practice a pro cess
may b e blocked if it is marked in Therefore a parallel pro cess can come out of the
blo cked status only if it is ready for service controlled by the lo cal scheduler and
the event out o ccurs controlled by the top level scheduler By letting only one
parallel pro cess b e marked out on each pro cessor at any time we can guarantee
that only one parallel pro cess time shares resources with sequential pro cesses in
each time slot
When a time slot is ended for the current parallel pro cess the servant will
move to a new no de The parallel pro cess asso ciated with that no de can then
b e serviced next However the movement of the servant is totally controlled
by an o ce manager which has a timer to determine when the servant is to
move and an algorithm to determine which no de the servant is to move to The
algorithm can b e simple ones such as the conventional round robin To obtain a
high system throughput however other more sophisticated scheduling schemes
may also b e considered The timer is to ensure that pro cesses can obtain their
(i) (i)
s in each scheduling round s or t allo cated service times that is t
s p
The use of registration o ces is similar to that of the two dimensional ma
trix adopted in the conventional coscheduling Each column of the matrix cor
resp onds to a time slot and each row to a pro cessor The coscheduling is then
controlled based on that matrix It is easy to see that the linked list on each
pro cessor plays the same role as a row of that matrix in coscheduling parallel
pro cesses However the key di erence is that our two level scheduling scheme
allows b oth parallel and sequential jobs to b e executed simultaneously
Lo ose Gang Scheduling
The conventional gang scheduler is centralised The system has a central con
troller At the end of each time slot the controller broadcasts a message to all
pro cessors The message contains the information ab out which parallel work
load will receive a service next The centralised system is easy to implement
esp ecially when the scheduling algorithm is simple However frequent signal
broadcasting for simultaneous context switch across the pro cessors may degrade
the overall system p erformance on machines such as networks of workstations
and space sharing p olicies may not easily b e adopted to enhance the e ciency
of resource utilisation Because in our system there is a registration o ce on
each pro cessor we can adopt a loose gang scheduling p olicy to alleviate these
problems
In our system there is a global job manager It is used to monitor the work
ing conditions of each pro cessor to lo cate and allo cate pro cessors and to assign
utilisation ratios to parallel jobs and to balance parallel and sequential work
loads We b elieve that resources in networks of workstations cannot e ciently
b e utilised without an e ective global job manager This global job manager is
also able to broadcast signals for the purp ose of synchronisation to co ordinate
the execution of parallel jobs However the signals need not b e frequently broad
cast for simultaneous context switch b etween time slots across the pro cessors
They are sent only once after each scheduling round or even many scheduling
rounds to adjust the p otential skew of the corresp onding time slots or simply
time skew across the pro cessors caused by using local job managers on each
pro cessor
There is a lo cal job manager on each pro cessor It is used to monitor and
rep ort to the global job manager the working conditions on that pro cessor It
also takes orders from the global job manager to prop erly set up its registration
o ce and to co ordinate the execution of parallel jobs with other pro cessors
With help of the global job manager the e ective coscheduling is guaranteed by
using lo cal job managers on each pro cessor
Time
T
pro cessor
6 1 2
Space
pro cessor
pro cessor
3 4
pro cessor
pro cessor
5
Fig The time space allo cation for six jobs on ve pro cessors
In the following we give a simple example which demonstrates more clearly
the e ectiveness of using the lo ose gang scheduling scheme and which also
presents another way of deriving the registration o ce for the scheme
Our simple example considers the execution of six jobs on ve pro cessors
We assume that the time space allo cation has already b een done that is the
numb er of pro cessor and the utilisation ratio have b een assigned for each job
as depicted in Fig For various reasons such as describ ed in the previous
sections the shap es of time space allo cation may not b e the same for each job as
indicated in the gure This will make it very di cult for a centralised controller
to coschedule jobs However the problem can easily b e solved by adopting our
lo ose gang scheduling
On each pro cessor we run a lo cal job manager and we also set up a schedul
ing table which is given by the global job manager Parallel pro cesses are then
scheduled according to this scheduling table In our example there are three dif
ferent scheduling tables as shown in Fig a The pro cesses and the lengths
of their allo cated time slots in a scheduling round are listed in each table in
an ordered manner It is easy to see that if the pro cessors are synchronised
at the b eginning of each scheduling round It is also p ossible that the pro ces
sors can b e synchronised once many scheduling rounds and lo cal job managers
schedule parallel pro cesses according to the given scheduling tables the correct
coscheduling across the pro cessors is then guaranteed
Because b oth content and size of each table vary from time to time during the
computation it is quite natural to implement the scheduling tables using linked
lists which results in our registration o ce A registration o ce on pro cessor
is depicted in Fig b Note each no de in the linked list has a p ointer which
p oints at the corresp onding pro cess so that any unnecessary search for parallel
pro cesses can b e avoided
pro cessor pro cessor
pro cessor
1 2 6 3 4 6 3 5
J J J J J J J J
1 2 6 3 4 6 3 5
a
pro cessor
H T
1 2 6
J J J
1 2 6
b
Fig a The scheduling tables assigned for each pro cessor and b The registration
o ce on pro cessor
With the collab oration of the global and lo cal job managers the system can
work correctly and e ectively A p otential disadvantage of the lo ose gang schedul
ing is that there is an additional cost for executing the coscheduling algorithm
(i)
(i)
or t are usually in on each pro cessor However in practice time slots t
p
order of seconds This extra cost for running a pro cess for coscheduling will b e
relatively very small
Conclusions
In this pap er we discussed some new ideas for e ectively scheduling b oth parallel
and sequential workloads on networks of workstations
To achieve a desired p erformance in a system with a variety of comp etitive
background workloads the key is to assign a sustained CPU utilisation ratio on
each pro cessor to a parallel job so that the p erformance b ecomes predictable
Because the resources in a system are limited however we cannot guarantee
that every job will b e given a sustained utilisation ratio One way to solve this
problem is that small parallel jobs are assigned a sustained ratio of resource
utilisation while each large parallel job is allo cated a large numb er of pro cessors
and assigned a utilisation ratio which can vary in a wide range according to
the current system workload Thus small jobs are not blo cked by larger ones
and a short turnaround time is guaranteed high e ciency of resource utilisation
can b e achieved and reasonably go o d p erformance for large jobs may also b e
obtained
To balance the workloads for b oth sequential and parallel pro cessing we
intro duced a two level scheduling scheme At the global level parallel jobs are
coscheduled so that they can obtain their prop er shares without interfering with
each other and they can also b e co ordinated across the pro cessors to achieve
high e ciency in parallel computation At the lo cal level many di erent p olicies
e g the busy waiting or spinning and the implicit coscheduling or two phase
blo cking can b e considered to schedule b oth parallel and sequential pro cesses
We intro duced a prop er share p olicy for e ectively scheduling pro cesses at the
lo cal level By adopting this p olicy we can obtain go o d p erformance for each
parallel job and also maintain go o d resp onse for interactive sequential jobs The
two level scheduling can b e implemented by adopting a registration o ce on
each pro cessor The organisation of the registration o ce which is also describ ed
in is simple and the main purp ose is to e ectively schedule parallel pro cesses
at b oth global and lo cal levels
We also intro duced a lo ose gang scheduling scheme to coschedule parallel jobs
across the pro cessors This scheme requires b oth global and lo cal job managers
The coscheduling is mainly controlled by lo cal job managers on each pro ces
sor so frequent signal broadcasting for simultaneous context switch across the
pro cessors is avoided There is only a bit extra work for global job manager to
adjust p otential time skew The name loose gang has two meanings First the
coscheduling is achieved by mainly using lo cal job managers but not just a cen
tral controller and second parallel pro cesses may time share their allo cated time
slots with sequential pro cesses Since b oth global and lo cal job managers play
e ective roles in job scheduling we think this may lead a way for us to nd
go o d strategies for e ciently scheduling b oth parallel and sequential workloads
on networks of workstations
A new system based on these ideas is currently under construction on a
distributed memory parallel machine the Fujitsu AP at the Australian
National University
References
T Agerwala J L Martin J H Mirza D C Sadler D M Dias and M Snir SP
system architecture IBM Systems Journal
S V Anastasiadis and K C Sevcik Parallel application scheduling on networks of
workstations Journal of Paral lel and Distributed Computing pp
T E Anderson D E Culler D A Patterson and the NOW team A case for
NOW networks of workstations IEEE Micro Feb pp
R H Arpaci A C Dusseau A M Vahdat L T Liu T E Anderson and D
A Patterson The interaction of parallel and sequential workloads on a network of
workstations Proceedings of ACM SIGMETRICS PERFORMANCE Joint
International Conference on Measurement and Modeling of Computer Systems
May pp
A C Arpaci Dusseau and D E Culler Extending prop ortional sha re scheduling
to a network of workstations Proceedings of International Conference on Paral lel
and Distributed Processing Techniques and Applications Las Vegas Nevada June
M Crovella P Das C Dubnicki T LeBlanc and E Markatos Multiprogramming
on multipro cessors Proceedings of the Third IEEE Symposium on Paral lel and
Distributed Processing Dec pp
J J Dongarra Performance of various computers using standard linear equations
software Technical Rep ort CS Computer Science Department University
of Tennessee Nov
A C Dusseau R H Arpaci and D E Culler E ective distributed scheduling of
parallel workloads Proceedings of ACM SIGMETRICS International Confer
ence
D G Feitelson and L Rudolph Gang scheduling p erformance b ene ts for ne
grained synchronisatio n Journal of Paral lel and Distributed Computing
Dec pp
D Ghosal G Serazzi and S K Tripathi The pro cessor working set and its use in
scheduling multipro cessor systems IEEE Transactions on Software Engineering
May pp
A Gupta A Tucker and S Urushibara The impact of op erating system schedul
ing p olicies and synchronisation metho ds on the p erformance of parallel applica
tions Proceedings of the ACM SIGMETRICS Conference on Measurement
and Modeling of Computer Systems May pp
K Li IVY A shared virtual memory system for parallel computing Proceedings
of International Conference on Paral lel Processing pp
S P Lo and V D Gligor A comparative analysis of multipro cessor scheduling
algorithms Proceedings of the th International Conference on Distributed Com
puting Systems Sept pp
V K Naik S K Setia and M S Squillante Performance analysis of job schedul
ing p olicies in parallel sup ercomputing environments Proceedings of Supercomput
ing Nov pp
V K Naik S K Setia and M S Squillante Pro cessor allo cation in multipro
grammed distributed memory parallel computer systems IBM Research Rep ort
RC
J K Ousterhout Scheduling techniques for concurrent systems Proceedings of
Third International Conference on Distributed Computing Systems May pp
E Rosti E Smirni L Dowdy G Serazzi and B M Carlson Robust partitioning
p olicies of multipro cessor systems Performance Evaluation pp
S K Setia M S Squillante and S K Tripathi Analysis of pro cessor allo cation in
multiprogrammed distributed memory parallel pro cessing systems IEEE Trans
actions on Paral lel and Distributed Systems April pp
I Stoica H Ab del wahab K Je ay S Baruah J Gehrke and C G Plaxton
A Prop ortional share resource allo cation algorithm for real time time shared sys
tems IEEE Real Time Systems Symposium Dec
K Suzaki H Tanuma S and Y Ichisugi Design of combination of time sharing
and space sharing for parallel task scheduling Proceedings of the International
Conference on Paral lel and Distributed Processing Techniques and Applications
Las Vegas Nevada Nov
C A Waldspurger and W E Weihl Stride scheduling deterministic prop ortional
share resource management Technical Rep ort MIT LCS TM MIT Lab ora
tory for Computer Science MIT June
J Zahorjan and E D Lazowska Spinning versus blo cking in parallel systems
with uncertainty Proceedings of the IFIP International Seminar on Performance
of Distributed and Paral lel Systems Dec pp
B B Zhou X Qu and R P Brent E ective scheduling in a mixed parallel and
sequential computing environment Proceedings of the th Euromicro Workshop on
Paral lel and Distributed Processing Madrid Jan
B B Zhou R P Brent D Walsh and K Suzaki A multi class time space sharing
system Tech Rep DCS and CSLab Australian National University in pro cess