Public Cluster : parallel machine with multi-block approach

Abstract

We introduce a new approach to enable an open and public parallel machine which is accessible for multi users with multi jobs belong to different blocks running at the same time. The concept is required especially for parallel machines which are dedicated for public use as implemented at the LIPI Public Cluster. We have deployed the simplest technique by running multi daemons of parallel processing engine with different configuration files specified for each user assigned to access the system, and also developed an integrated system to fully control and monitor the whole system over web. A brief performance analysis is also given for Message Parsing Interface (MPI) engine. It is shown that the proposed approach is quite reliable and affect the whole performances only slightly.

Full text

Page 1

Public Cluster : parallel machine with multi-block approach
Z. Akbar1*, Slamet2, B.I. Ajinagoro3, G.J. Ohara3, I. Firmansyah1, B. Hermanto1, L.T. Handoko1
1 Group for Theoretical and Computational Physics, Research Center for Physics, Indonesian Institute of Sciences, Kompleks
Puspiptek Serpong, Tangerang 15310, Indonesia
2 Faculty of Computer Science, University of Indonesia, Kampus UI Depok, Depok 16424, Indonesia
3Faculty of Electronics Engineering, STT Telkom, Jl. Telekomunikasi, Dayeuh Kolot, Bandung 40257, Indonesia
We introduce a new approach to enable an open and public parallel machine which is accessible for multi users
with multi jobs belong to different blocks running at the same time. The concept is required especially for parallel
machines which are dedicated for public use as implemented at the LIPI Public Cluster. We have deployed the
simplest technique by running multi daemons of parallel processing engine with different configuration files
specified for each user assigned to access the system, and also developed an integrated system to fully controll and
monitor the whole system over web. A brief performance analysis is also given for Message Parsing Interface (MPI)
engine. It is shown that the proposed approach is quite reliable and affect the whole performances only slightly.
1. Introduction
Along with the advances of scientific researches,
especially in the field of basic natural sciences in the last
decades, the needs on advanced computing is increasing
exponentially. Such kind of advanced computings require
higher specifcations on hardwares which leads then to
astronomical cost to realize. One key solution for this
problem is the parallel / cluster machine.
Nowadays, clustering (the low specs and low cost
machines) becomes the mainstream to realize an advanced
computing system comparable to or in most cases better than
the conventional mainframe based system with significant
reduced cost [1]. Generally the cluster is designed to
perform a single (huge) computational task at certain period.
This makes the cluster system is in general exclusive and not
at the level of appropriate cost for most potential users,
neither the young beginners nor the small research group,
especially in the developing countries.
It is clear that the cluster is in that sense still costy,
although there is certainly needs to educate the young
generation to be the next users familiar with parallel
programmings etc. This background motivates us to develop
an open and free cluster system for public, namely the LIPI
Public Cluster (LPC) [2,3,4].
Concerning those characteristics, the public cluster
should be accessible and user-friendly for all users with
various level of knowledges on parallel programming and
also various ways of accessing the system in any platforms.
This can be achieved by deploying the web-based interfaces
in all aspects. Therefore, the main issues are then the
security from anonymous users, avoiding the interferences
among different tasks running on multi blocks of the nodes
and the real-time monitoing and control over web for both
administrators and users.
In this paper we present the simplest method to fill such
requirements. In the subsequent section we first introduce
the multi-block system, followed by our concept on public
cluster. Finally we provide a brief performance test on the
current system before coming to the conclusion.
2. Multi-block system
In order to run a particular task on several machines, we
must implement an interface for the process management
and communication among the nodes. In ur system we
deploy the popular one, that is the Message-Passing
Interface (MPI) [5]. MPI is widely used due to its portability
to run MPI in any platforms through the message-passing
protocol. Further, we have also implemeted the MPICH2
that is the upgraded version of the MPI developed by the
Argonne National Laboratory [6]. InMPICH2 the process
management and the communication is completely splitted
off. So, the initial runtime environment contains several
daemons called MPD's which each of them has a task to
initiate the communication among the nodes before running
the main programme. This mechanism is crucial since it
enables us to distinguish the errors either in the process or
communication.
Each MPD daemon is configurable through its
configuration file or optional parameters in command line.
Since in our system each user has their own configuration
and daemon of MPD, MPICH suits well our needs on a
cluster system with different parallel computation at the
same time. Further, each user who is assigned to use the
cluster with partiicular configuration and daemon is called as
a block. Therefore, we define the multi-block as a case
where several users with different blocks access the system
at the same time.
MPD is a process management assembles daemons
which execute the MPI modules. In order to form these
daemons as a ring, there should be a node works as a master.
This master node will boot the MPD in each node member.
This flow is depicted in Fig. 1.
Fig. 1. The multi-block system at the LPC.

Page 2

Fig. 1 shows logically the creation of two rings of MPD.
The first block (a) contains three nodes in addition to one
MPD in the master node. On the other hand, the second
block (b) has two nodes and also one MPD in the master
node. Each node is restricted to have only one running
MPD, while the master node is allowed to run multiple
MPD's as long as each MPD belongs to different user. Under
this mechanism the interferences among the nodes could be
completely avoided.
The mechanism to enable multi-block system requires
several procedures :
●
Each node, including the master node, communicates
each other using passwordless access, for instance ssh
with RSA-key. This can be accomplished easily without
copying the authorized keys to all assigned nodes, for
example by exporting the user's home directory through
NFS.
●
The node names which form the ring of MPD are
located at the file mpd.hosts in the user's home
directory. In the LPC, the node's names are assigned by
the administrator and unchangeable by the users.
●
The main configuration file for MPD, .mpd.conf, is
located in the user's home directory. The important
parameter is MPD_SECRETWORD=<password>
which is crucial for security of the MPD's ring. The
<password> should be the same for all nodes in the
same ring.
●
The master node behaves as dual-homed, that is it has
multi network interfaces (NIC's) and IP addresses.
Because we must split the NIC's off for the external and
the internal accesses. The parameter to assign the
appropriate NIC for the MPDdaemon is --ifhn
=<ip_address>.
●
It should be kept in mind that without any restriction,
the number of processes running in a node is infinity
and the tasks will be distributed in rond-robin way. This
can be configured through the parameter
<node_name>:<process_number>.
3. Public cluster
After discussing the multi-block system in the
preceeding section, now let us consider the concept of public
cluster.
First of all, let us recall the differences between the LPC
and the other clusters in general :
●
The users are all anonymous, i.e. everyone can register
themself and then start using the cluster after getting
approved and assigned with particular number of nodes
by the administrator.
The specification and the node number is fully
determined by the administrator. So, the administrator
has full control on the system, and then can shut unused
nodes down if necessary to save the resources. This is
crucial since the LPC consists of nodes with several
specifications ranging from P4 with huge memories and
spaces to the old 486 PC's. This means the administrator
is able to assign an appropriate block of nodes
concerning the level of computation will be perforned.
In this sense, the distribution mechanism here is
completely different with the known openPBS [7].
The system is fully accessible, monitorable and
controllable through the web interface by both
administrator and users without having the ssh access.
This would also improve the security concerning
anonymous users accessing the system.
The work-flow of LPC is diagrammatically depicted in
Fig. 2. The flow can be explained as below :
1. A new user does the registration by providing the
personal data, the content of job will be performed and
the number of nodes requested for the job.
2.The application is reviewed and verified by the
administrator. The administrator will also assign the
nodes provided for the approved user and the usage
period.
3.Reconfirmation by user. If the user agrees with the
provided nodes and the usage period, the administrator
will switch the nodes on and activate all daemons.
4. The user should adjust their parallel programme to fit
the provided nodes.
5. The user uploads all necessary programme and libraries
if any. At this stage the programme can be executed
immediately.
6.The administrator and automated system will monitor
the usage of all running users.
7. Finished job can be downloaded by the user.
●
●
Fig. 3. Mesurement of bisection bandwith in single block
(red line) and two blocks (green line) clusters.
Fig. 2. The interaction flow of administrator and users at
the LPC.
1
USER
Public
Cluste
r
ADMI
N
2
4
5
3
6
78

Page 3

8. Once the usage period is over, the nodes are shutted
down automatically.
4. Performance test
Now, we are ready to study the performance of the
system. Our concern is of course how far the multi-block
approach influences the whole performance ? This point is
crucial, especially since all blocks (rings) rely on the same
master node where it also acts as a node in each block.
The benchmark test is done by dividing the whole system to be
two uncomparable blocks, i.e. there is a significant difference of
node numbers between both blocks. As a benchmark tool, we use
mpptest from the Argonne National Laboratory [8] and measure
the bisection bandwith.
The benchmark is performed by running one block while the
other is shutted down. Thereafter, both blocks are activated
simultaneously. The result is shown in Fig. 3, where the red line is
for single block, and the green line is for both blocks.
Form the figure, it is clear that running several blocks at the
same master node is still reliable and affect the whole
performances only slightly. Of course, this performance is not
guaranteed for all kind of parallel programmings, but at least for
certain objectives at the LPC this result proves and justifies the
current architecture of LPC.
5. Conclusion
We have introduced an approach to enable multi-block
system in a cluster dedicated for public use. Equipping the
cluster with an integrated control and monitoring system,
and also the interactive web interface would realize a low
cost and easy-to-use or -maintain parallel machine for
public. This would be a significant contribution to encourage
and educate young generation on parallel programming
which should be crucial for advanced researches in any areas
in the future.
Acknowledgment
Slamet greatly appreciates the support from his
colleagues at the Faculty of Computer Science, University of
Indonesia. B.I. Ajinagoro and G.J. Ohara appreciate their
colleagues at the Faculty of Engineering, STT Telkom. This
work is financially supported by the Riset Kompetitif LIPI
in fiscal year 2007.
References
(1) A. Goscinski, M. Hobbs and J. Silcock: Cluster Computing, Vol. 4, pp.
145–156 (2001).
(2) LIPI Public Cluster, http://www.cluster.lipi.go.id.
(3) L.T. Handoko,, Indonesian Copyright, No. B 268487 (2007).
(4) Z. Akbar, Slamet, B.I. Ajinagoro, G.J. Ohara, I. Firmansyah, B.
Hermanto and L.T. Handoko: Indonesian Patent, under registration
(2007)
(5) Message Passing Interface Forum, International Journal of High
Performance Computing Applications, pp. 1–299 (1998).
(6) William Gropp, Learning from the Success of MPI, Argonne National
Laboratory (2001).
(7) OpenPBS, http://www.openpbs.org.
(8) William Gropp and Ewing Lusk, the Proceedings of PVMMPI'99
(1999).