 |
SCore Glossary |
A
B C D
E F G
H I J
K L M
N O P
Q R S
T U V
W X Y
Z
- AFS
- A distributed file system originally developed at CMU. AFS is a
product of IBM Transarc Lab
See
http://www.transarc.com/index.html.
- (PM) Agent
- PM/Agent is a
PM device implemented with a
agent (daemon) process which processes low-level protocol, for
example, maintain reliability. Currently, an agent using UDP
protocol is supported. PM/Agent is not a
user-level communication.
- auto-negotiation
- Ethernet has several physical layers in its speed (10Mbps,
100Mbps and 1Gbps) and direction (uni- or bi-direction). Network
hardware has to negotiate with each other how to link with at the
very first connection. Most Ethernet hardware
(NIC or HUB) has
auto-negotiation facility, however, sometimes network hardware
fails to negotiate, and a connection is made in
mistakenly. When this happens, communication performance can be
degraded far from its specifications. To avoid this situation, user
has to set up the connection manually.
A
B
C D E
F G H
I J K
L M N
O P Q
R S T
U V W
X Y Z
- binary type
- A unique symbol to distinguish processor type and host OS.
A B
C
D E F
G H I
J K L
M N O
P Q R
S T U
V W X
Y Z
- C-Area
- C-Area is a System-V shared memory region between
SCore-D process and user process in
a user parallel
process. When a user process is forked, all the information to
initialize user process is passed via C-Area, such as node
number, number of nodes, PMv2 context information, and so
on. C-Area is also used when a user program raises a
systemcall to SCore-D.
- (PM) channel
- PM channel is a virtualized network. In PMv2,
physical network is multiplexed by software. Each multiplexed
network is called (PM) channel.
- cluster
- Networked computers. In SCore, a cluster consists of
homogeneous computers.
- cluster of clusters
- Networked clusters.
In SCore, a cluster of clusters can be heterogeneous.
- compute host
- A host that is a member of a
cluster.
- cluster host
- See compute host.
- (PM) Composite
- PM/Composite device is not an actual network device. It
only forwards communication requests to other real PM
device(s) according to a routing table which
PM/Composite has. Having unique PM/Composite,
users can combinme any arbitray
PM devices.
PM/Composite is also designed
for thread safe, although the other real PM devices are
not. When communication requests are forwarded,
PM/Composite locks not to accept the other request from the
other thread. Therfore,
SCore-D always allocates one
PM/Composite device and the othe PM device(s) for
a user parallel process running on two or more nodes.
- (PM) configuration file
- A file needed to configure a
PM device as a network. The file
format of the configuraiton file depends on a
PM device.
- (PM) context
- A set of state representing PM channel status.
- co-scheduling
- A scheduling technique for multi processors proposed by
Ousterhout 1982. In co-scheduling, context switching of some or all
processes in a parallel process is synchronized to obtain
higher efficiency.
See also gang-scheduling.
- COW
- COW stands for Cluster Of Workstations.
A B C
D
E F G
H I J
K L M
N O P
Q R S
T U V
W X Y
Z
- deadlock detection
- Deadlock detection is one of the most unique features of
SCore-D.
SCore-D can detect
a deadlock situation of a user parallel process.
SCore-D can detect more than deadlock. The
situation such that all process in a parallel process is just
waiting for incoming messages can also be detected. Strictly
speaking, this kind of situation is not a deadlock, since there is
no resource contention. The term of 'deadlock' is used here, just
because it inspires users the situation where parallel computation
will never proceed. This situation is defined as
that there is no transient messages and no busy processes in a
parallel process.
SCore-D checks
this everytime multiple context switch
(gang scheduling) takes place.
- (PM) device
-
There are several PM devices having the same PMv2
API. Currently, there are six devices,
PM/Etherner,
PM/Myrinet,
PM/Shmem,
PM/Agent,
PM/RHiNET, and
PM/Composite developed.
- (SCore-D) device
-
Actually SCore-D device is not a physical device, but a
sequential Unix process written in C. Thus users can develop their
own devices, and attach them to
SCore-D. This
SCore-D device is one of the unique features of
SCore-D. Using SCore-D device, user parallel
program can wait for the I/O through the device in a blocking
way. One user of SCore system developed a parallel server
running on a cluster accepting requests from WWW. The user also
developed an SCore-D device to accept requests from a CGI
program, and the server is waiting for the requests. When there is
no request, the server parallel process is blocked. When a
request arrives, server parallel process begins to run with
other users parallel procceses in
TSSS. Thus, the cluster can be used for
a server and program develepment simultaneously.
- DMA
- DMA stands for Direct Memory Access. A memory copying hardware
to move data between memory regions, or an I/O device and a memory
region. In many cases, DMA is used for moving I/O data. In general,
DMA has longer setup time than CPU memory copying, but DMA
may have larger copying bandwidth.
A B C
D
E
F G H
I J K
L M N
O P Q
R S T
U V W
X Y Z
- (PM) Ethernet
- PM/Ethernet is a
PM device for Ethernet
NICs. It supports any 10, 100 and 1Gbps
NICs. For enabling high
performance and high portability, PM/Ethernet is desinged
in a unique way. There is a small kernel patch to discriminate IP
(Internet Protocol) packets and PM/Ethernet
packets. IP packets are handled in IP protocol layer, and
PM/Ethernet packets are handled in the PM/Ethernet
driver. Thus, an Ethernet NIC can handle both
IP and PM/Ehernet packets. PM/Ethernet achieves
higher performance compared with the communication using UDP or TCP
protocol, however, there is one drawback. Further, PM/Ethernet supports
network
trunking. PM/Ethernet uses
the Ethernet MAC address directly, and there is no way to route the
PM/Ethernet packets out of a network segment. In this case,
use PM/Agent
instead. PM/Ethernet is not a
user-level communication.
- EIT
- EIT stands for Easy Installation Tool to install SCore
Cluster System easily with a GUI. It is designed for those
who are unfamiliar with Linux installation.
A B C
D E
F
G H I
J K L
M N O
P Q R
S T U
V W X
Y Z
- FEP
- FEP stands for Front End Process.
scrun is a FEP.
- FQDN
- FQDN stands for Full Qualified Domain Name. For example,
pccluster.org is a FQDN. See also official name.
A B C
D E F
G
H I J
K L M
N O P
Q R S
T U V
W X Y
Z
- global clock
- PMv2 has some API functions for supporting globaly
synchronized clock. However, these functions are only effective for
PM device having a
clock. Currently only
Myrinet NIC has
such a clock.
SCore-D cooperating with PMv2
periodically synchronize NIC clocks. So that user
program can read and use the clock value. However, due to the
distributed nature, each clock is not synchronized fully, but
consistently. This means that the there can be some clock skew, but
the skew is not observed. Further, the clock synchronization is done
in a best-effort way, and
SCore-D does not guarantee for the
consistency.
- gang scheduling
- A time sharing scheduling technique in that all process
switching in a parallel process is synchronized. In this
way, a time sharing scheduling can be practical even for fine grain
parallel programs. In
SCore-D, user parallel
processes are gang-scheduled.
See also co-scheduling.
- global time stamp
- See global clock.
A B C
D E F
G
H
I J K
L M N
O P Q
R S T
U V W
X Y Z
- host
- A computer having unique hostname(s).
A B C
D E F
G H
I
J K L
M N O
P Q R
S T U
V W X
Y Z
- idle flag
- It is difficult to distinguish idle or busy state of a process
waiting for incoming messages with polling. This is the typical case
with the parallel programs using
user-level communication.
The idle flag
in C-Area is used for
SCore-D
to obtain the status of a user process. Parallel runtime libraries,
such as
MPICH-SCore or
MPC++/
ULT set the flag
appropriately. SCore-D uses this
information to determine if the user parallel process is
deadlocked, and output the user process activity via
loadmonitor.
A B C
D E F
G H I
J K
L
M N O
P Q R
S T U
V W X
Y Z
- LANai (processor)
- A special purpose processor for controling Myrinet
NIC developed by
Myricom.
- Linux
- The most populer Unix like operating system developed by Linus
Torvalds. Several companies (such
as RedHat,
Suse and
Turbo Linux) are distributing Linux as a
software package.
See http://www.linux.org/.
- Loadmonitor (3D)
- SCore-D samples the
idle flag periodically (200ms, in
default configuration), and
its status is output graphically via loadmonitor program
running on FEP host. Currently
SCore-D samples two
kinds of state on ech node, one is the idle flag and the
other is PMv2 messaging status.
A B C
D E F
G H I
J K L
M
N O P
Q R S
T U V
W X Y
Z
- Mesa
- A 3-D graphics library supporting
OpenGL like API.
See http://www.mesa3d.org/.
- Messageboard
- Messageboard is a server to lock
cluster hosts.
- MTTL
- MTTL stands for Multi-Thread Template Library. It is a
C++ template library for multi-thread programming on distributed
memory clusters. MTTL is a part of
MPC++.
- MPC++
- A multi-threaded C++. MPC++ stands for Multi-Thread,
Meta-Programming or Massively Parallel C++. MPC++ has two
unique features. One is the multi-thread feature which is
implemented using C++ template, called
MTTL. And the other
is meta-level architecture. MPC++ meta-level-architecture
is still under development, due to the complexity of the C++
syntax.
- MPI
- MPI stands for Message Passing Interface.
See
http://www-unix.mcs.anl.gov/mpi/.
- MPICH
- A reference implementation of MPI developed by Argonne National
Laboratory. See
http://www-unix.mcs.anl.gov/mpi/mpich/.
- MPICH-SCore
- An MPICH porting for SCore Cluster System Software.
- Myricom
- See http://www.myri.com/.
- Myrinet
- A gigabit-per-second class network developed by Myricom.
- (PM) Myrinet
- PM/Myrinet is a
PM device using
Myrinet NIC.
Since PM/Myrinet was the first
PM device among the others,
PM/Myrinet is tuned very well, stable, and providing the
largest optional operations. PM/Myrinet is a
user-level communication.
A B C
D E F
G H I
J K L
M
N
O P Q
R S T
U V W
X Y Z
- network preemption
- In user-level communication, user
programs are allowed to access hardware to obtain
higher performance. In a multi-programming (multi-user) environment,
however, accessing hardware directly should be prohibited, or,
hardware status must be saved and restored when switching
processes.
SCore-D and
PMv2 cooperate for the latter
processing. This procedure is called network preemption in
SCore.
- network interleaving
- SCore-D supports interleaving two or more
networks to obtain higher bisection bandwidth.
- network trunking
- See Ethernet trunking.
- NFS
- The most widely used distributed file system.
- NIC
- NIC stands for Network Interface Card. In general, NIC is a
interface board plugged in an I/O bus (e.g. PCI) slot.
- node
- An entity to communicate with. A node is distinguished with the
others by a node number (a positive integer).
In MPI, the node number is called "rank."
- NOW
- NOW stands for Network Of Workstations.
A B C
D E F
G H I
J K L
M N
O
P Q R
S T U
V W X
Y Z
- official (host)
name
- An official name is a hostname obtained by the
gethostbyname function in glibc.
For more details, see the gethostbyname(3) man page.
- Omni
- An OpenMP compiler developed
at RWCP. Omni supports not only SMP parallel
programming environment, but also distributed memory parallel
programming environment.
- OpenGL
- OpenGL is a widely used 3-D graphics library.
See http://www.opengl.org/.
- OpenMP
- OpenMP is an SMP programming environment.
See http://www.openmp.org/.
A B C
D E F
G H I
J K L
M N O
P
Q R S
T U V
W X Y
Z
- page coloring
- If a processor has a set-associative cache, memory pages having
the same address bit pattern are cached in the same cache
entry, and if the entry overflows then cache miss ratio
increases. Page coloring is a technique for avoiding this
situation. Memory pages are colored according to the bit pattern,
and allocate different colored pages if possible. In SCore
Cluster System Software, a kernel patch for page coloring is
included.
- parallel job
- In SCore Cluster System Software, a parallel job is a
set of parallel processes. However, in the
parallel job scheduling research community, a parallel job means a
set of processes doing a parallel computation.
- parallel process
- A parallel process is a set of Unix processes communicating each
other for a parallel computation.
- PBS
- PBS stands for Protable Batch System. PBS is a bacth job
scheduling system, originally developed by NASA.
See http://pbs.mrj.com.
- PCC
- PCC stands for PC Cluster.
- pin-down cache
- PMv2's pin-down cache is a kind of cache holding
locked mmemory pages to reduce the overhead of memory page locking.
See also
remote memory access.
- PM
- A low level communication library and a device driver dedicated
for Myrinet. PM was used in
SCore Cluster System Software Version 2.X.
See also PMv2.
- PMv2[pee-em-vee-two]
- PMv2 is the name of the API for low-level, high-performance
communication library supporintg various network
devices. PMv2 is the successor of
PM. PM stands for nothing. There is a
joke for the origin of PM. "PM library was written only in the
afternoon." See also PM device.
- PM/Agent
- See (PM) Agent.
- PM/Composite
- See (PM) Composite.
- PM/Ethernet
- See (PM) Ethernet.
- PM/Myrinet
- See (PM) Myrinet.
- PM/RHiNet
- See (PM) RHiNET.
- PM/Shmem
- See (PM) Shmem.
- PM/UDP
- From SCore 3.2, PM/UDP is implemented as a PM/Agent.
See (PM) Agent.
A B C
D E F
G H I
J K L
M N O
P
Q
R
S T U
V W X
Y Z
- RedHat Linux
- A Linux distribution.
See http://www.redhat.com/.
- (PM) RHiNET
- RHiNET stands for RWCP Hight-performance NETwork.
RHiNET is a high-performance network developed by RWCP.
- (PM) remote memory access
- PMv2 supprts several ways for
zero-copy communication. In
PMv2, the communication using zero-copy communication
technique is called remote memory access. Since zero-copy
communication uses NIC DMA,
memory pages must belocked before DMA, so that
the physical address of the memory
region sit on a same logical address. Because most of modern OS,
such as Linux, swaps memory pages
between physical memory and disk. The locking and unlocking memory
pages must be done by an OS, and they are costly operations. To reduce
the memory locking overhead, PMv2 has a kind of cache for
locked memory pages, called
pin-down cache.
- RPM
- RPM stands for Red hat Package Manager. See
rpm(8) Unix man page or
http://www.rpm.org/.
- RWCP
- RWCP stands for Real World Computing Partnership (Project). RWCP is
a ten year project funded by Ministry of International Trade and
Industry (MITI) of Japan, starting from 1992.
In despite of the name, none of the RWCP researchers uses
FORTRAN nor COBOL, and wears tie in their office
(See also
Real World in
The New Hacker's Dictionary).
Actually the name in Japanese means "New Information Processing
Technology", however, it is never used because its abbreviation is
"NIPT."
A B C
D E F
G H I
J K L
M N O
P Q R
S
T U V
W X Y
Z
- SCASH [es-cache]
- Yet another software distributed shared memory system in
SCore Cluster System. SCASH is also a runtime
library for Omni enabling shared
memory programming on distributed memory clusters.
- Scoop
- A tool for monitoring
cluster hosts.
- SCore [es-core]
- SCore stands for System CORE. At first, SCore was the
name of parallel operating system, but now it is the name of a
cluster system software.
- SCore-D
[es-core-dee]
- The name of SCore-D used be a user-level (daemon)
implementation of a parallel operating system. Now it is the name of
a cluster operating system.
- scored (scout command)
- scored scout command is to
run SCore-D in a multi-user mode.
- Scoreboard
- A database program holding all the information to form a cluster
or a cluster of clusters.
- scorehosts.db
- The default database file name for scoreboard.
- Scout
- Scout provides a remote shell execution environment in
which Unix commands are executed in SIMD parallel way or
SCore-D runs.
- scout command
- Scout command is a special command effective only
within a scout environment. Currently,
scout, scrun, and
scored are the scout
commands. When scout is invoked outside of a scout
environment, it is a normal user command to create
a scout environment.
When scout is invoked
within a scout environment as a scout command, it
checks if the current scout environment is correctly
operating or not.
- scout environment
- Scout environment is a virtualized (sub)cluster
created by the scout command. In a scout
environment, users can use local host and
cluster hosts in a
transparent way. Not to interference with the other users,
scout locks
cluster hosts via
messageboard.
- scrun (command)
- scrun command is to run a user parallel program on a
cluster. When scrun command is invoked inside
of a scout environment without
scored option, as a
scout command, it executes
user parallel program on
SCore-D running in a single-user
mode. When scrun command is invoked with scored
option, regardless of the scout environment where
scrun is invoked, scrun submit a parallel job to
the SCore-D specified with the
scored option running in a multi-user mode. In both
cases, the invoked scrun process is called a
FEP process.
- server host
- A host serving for a
cluster. Usually, the
messageboard
server, and the scoreboard server
are running on the server host.
- shadow process
- In general, when a user process terminates, the memory pages for
the process are reclaimed. Suppose that a user program is using
zero-copy communication, and a user
process terminates while NIC is copying packets
into the process's address space via DMA. The
memory pages are reclaimed and reused, but NIC
DMA is still
working. Eventually reused memory contents might be clobbered by the
DMA. In SCore, to avoid this situation,
SCore runtime forks a process via Linux
clone(2) systemcall,
so that the memory pages are not reclaimed until
entire parallel process terminates. This cloned process is called
shadow process.
- (PM) Shmem
- PM/Shmem is a
PM device. It does not use
any NIC, but it uses a memory-mapped segement
shared by the processes in a SMP host to communicate with the
other process. There is a special Unix
device associated with PM/Shmem for enabling less-copy
communication to get higher communication performance.
- SMP
- SMP stands for Symmetric Multi-Processor. In
scorehosts.db file of
SCore, each host record must
have a smp attribute. The value of the smp
attribute must be the number of processor in a
host.
- SuSE Linux
- A Linux distribution.
See http://www.suse.com/.
- (SCore-D) systemcall
- As most operating systems have systemcalls,
SCore-D
cluster operating system has several systemcalls.
A B C
D E F
G H I
J K L
M N O
P Q R
S
T
U V W
X Y Z
- Time Slice
- An interval of process swithcing.
In SCore-D, time slice is an
interval of gang scheduling.
- (Ethernet) Trunking
- (Ethernet) trunking is a technology for combining two
or more Ethernet to get higher bandwidth.
PM/Etherner supprts (Ethernet)
trunking.
- TSSS
- TSSS stands for Time Space Sharing
Scheduling.
SCore-D packs parallel jobs
in a binary buddy way, and the jobs are scheduled in space sharing
and time sharing (gang scheduling)
simultaneously.
- Turbo Linux
- A Linux distribution. See
http://www.turbolinux.com/.
A B C
D E F
G H I
J K L
M N O
P Q R
S T
U
V W X
Y Z
- ULT
- ULT stands for User-Level Thread or Ultra Light
Thread. ULT is a multi-thread runtime library for
MTTL.
SCore-D is written in
MTTL using ULT.
- user-level communication
- Nowadays, NICs are getting faster and faster,
and communicatin latency can be less than 10
micro-seconds. And communication software becomes the bottleneck.
User-level communication is developed as a technique to reduce the
software overhead. In user-level communication, user program is
allowed to access NIC hardware directly, not by
some systemcalls. Incoming messages are polled, and interrupts are
disabled. In this way, neither systemcalls nor interrupts are
used in user-level communication, because they are costly operation
in recent high speed RISC processors.
A B C
D E F
G H I
J K L
M N O
P Q R
S T U
V W
X Y Z
- WSC
- WSC stands for WorkStation Cluster.
A B C
D E F
G H I
J K L
M N O
P Q R
S T U
V W
X Y
Z
- zero-copy communication
- Copying a memory region in an I/O used be a costly operation. To
get higher communication performance, copying has been considered
harmful. To avoid message copying, DMA of a NIC is used. This technique is
called zero-copy communication. In PMv2,
several zero-copy communication functions are supported as
remote memory access
functions. However, this technique is not
always effective. Some processors nowadays have higher bandwidth
than DMA. Further, setup time of DMA engine is
not negligible. It should be noted that zero-copy communication may
not contribute to the performance of applications.
- (SCore) zombie process
- A process of a user parallel process keeps alive even when a
SCore-D terminates. This situation
may happen when SCore-D crashes for some reasons. In Unix,
zombie process almost nothing harmful. SCore zombie
process, however, just like a zombie in a horror movie does,
sabotages SCore-D, because it keeps
holding some PM device.
- CREDIT
- This document is a part of the SCore cluster system software
developed at PC Cluster Consortium, Japan.
Copyright (C) 2003-2004 PC Cluster Consortium.
