NASA Technical Reports Server (NTRS) 19930005858: Storage needs in future supercomputer environments

20 Pages·1.1 MB·English

by  NASA Technical Reports Server (NTRS)

, N 9 - 15 047
Storage Needs in Future
Supercomputer Environments
Notes for the presentation by:
Sam Coleman
Lawrence Livermore National Laboratory
July 25, 1991 at the
NASA Goddard "Mass Storage Workshop"
Introduction the devices currently attached to them are
individually slow; storage systems based on
The Lawrence Livermore National this architecture will become bottlenecks
Laboratory (LLNL) is a Department of on HIPPI and other high-performance
Energy contractor, managed by the networks. Computers with the I/O-channel
University of California since 1952. Major performance to match these networks will
projects at the Laboratory include the be even more expensive than the current
machines.
Strategic Defense Initiative, nuclear
weapon design, magnetic and laser fusion,
The need for higher-performance storage
laser isotope separation and weather mod-
eling. The Laboratory employs about 8,000 systems is being driven by the remarkable
peopJe. There are two major computer advances in processor and memory tech-
centers: The Livermore Computer Center nology available on relatively inexpensive
and the National Energy Research workstations; the same technology is
Supercomputer Center. making high-performance networks pos-
sible. These advances will encourage
scientific-visualization projects and other
As we increase the computing capacity of
applications capable of generating and
LLNL systems and develop new applica-
tions, the need for archival capacity will absorbing quantities of data that can only be
increase. Rather than quantify that in- imagined today.
crease, I will discuss the hardware and
software architectures that we will need to To provide cost-effective, high-perfor-
mance storage, we need an architecture like
support advanced applications.
that shown in Figure 1. In this example,
striped storage devices, connected to a
Storage Architectures
HIPPI network through device controllers,
transmit large blocks of data at high speed.
The architecture of traditional super-
Storage system clients send requests over a
computer centers, like those at Livermore,
lower-performance network, like an
include host machines and storage systems
Ethernet, to a workstation-class machine
linked by a network. Storage nodes consist
controlling the storage system. This
of storage devices connected to computers
machine directs the device controllers, also
that manage those devices. These comput-
over a lower-performance path, to send
ers, usually large Amdahl or IBM main-
data to or from the HIPPI network. Control
frames, are expensive because they include
messages could also be directed over the
many I/0 channels for high aggregate
HIPPI network, but these small messages
performance. However, these channels and
276
would decrease the efficiency of moving formance of the system when large data
large data blocks; since control messages blocks are accessed (this architecture will
are small, sending them over a slower not be efficient for applications, like NFS,
network will not degrade the overall per- that transmit small data blocks).
Workstations
Controller
Striped
Devices
Workstation
Class Machine
Lower-Performance Control Network
A High-Performance Storage Architecture
Figure 1
To make the architecture in Figure 1 High-capacity media, with at least
efficient, we will need the following the capacity of the largest D2 tape
components: cartridges;
Programmable device controllers • Robotics to mount volumes quickly;
imbedding relatively high-level
data-transfer protocols; Devices and systems that are more
reliable than the 1-in-10 '2 error
High-performance, possibly striped, rates quoted today; and
archival storage devices to match the
performance of the HIPPI network. Devices that are less expensive than
These devices should be faster than the current high-performance de-
the D1 and D2 magnetic tapes being vices.
developed today;
277
In short, we need reliable, automated As a step toward the Figure 1 architecture,
archival devices with the capacity of Creo we are investigating the architecture
optical tapes (one terabyte per reel), the shown in Figure 2; we will connect existing
performance of Maximum Strategy disks storage devices to our Network Systems
(tens of megabytes per second), and the Corp. HYPERchannel, controlled by a
cost of 8mm tape cartridge systems (less workstation-based UniTree System. Even
than $100,000). though the hardW_ire connections are
Control
Crays Data
E
t
h Mover Gray Disks
e
r
n HYPERchannel
e
t
s
UniTree DX
Workstation STK 3480 Tapes
Amdahl 6380 Disks
An Interim Storage Architecture at LLNL
Figure 2
available today, the necessary software is mainframes that we use to control the
not. In particular, there is no high-level current archive.
file-transport software in the NSC DX
HYPERchannel adapter. As an interim Software Needs
solution, we will put IEEE movers _on our
host machines, allowing direct file-
To implement high.performance storage
transport to and from the storage devices
architectures, we need file-transport
over the HYPERchannel. The UniTree
software that supports the network-
workstation will provide service to client
attached devices in Figure 1. Whether or
workstations and other network machines.
not the TCP/IP and OSI protocols can
This is acceptable, in the near term,
transmit data at high speeds is subject to
because most of the archival load comes
debate; if not, we will have to develop new
from the larger host machines. This
protocols.
architecture will replace the =Amdahl
278
From the human client's point of view, we Replication
need software systems that provide Clients do not know if objects or
transparent access to storage. Several services are replicated, and services
transparencies are described in the IEEE do not know if clients are replicated.
Mass Storage System Reference Model
document: 1 Semantic
The behavior of operations is inde-
Access pendent of the location of operands
Clients do not know if objects or and the type of failures that occur.
services are local or remote.
Syntactic
Concurrency Clients use the same operations and
Clients are not aware that other parameters to access local and re-
clients are using services concur- mote objects and services.
rently.
The IEEE Reference Model
Data representation
Clients are not aware that different One way to achieve transparency is to de-
data representations are used in velop distributed storage systems that span
different parts of the system. clients environments. In homogeneous
environments, like clusters of Digital
Execution Equipment Corp. machines, transparency
Programs can execute in any location can be achieved using proprietary soft-
without being changed. ware. In more heterogeneous super-
computer centers, standard software,
Fault running on a variety of machines, is
Clients are not aware that certain needed. The IEEE Storage System Standards
faults have occurred. Working Group is developing standards
(project 1244) on which transparent
Identity software can be built. These standards will
Services do not make use of the be based on the reference model shown in
identity of their clients. Figure 3. The modules in the model are:
Location Application
Clients do not know where objects or Normal client applications codes.
services are located.
Bitfile Client
Migration This module represents the library
Clients are not aware that services routines or the system calls that in-
have moved. terface the application to the Bitfile
Server, the Name Server, and the
Naming Mover.
Objects have globally unique names
which are independent of resource Bltfile Server
and accessor location. The Bitfile Server manages abstract
objects called bitfiles that represent
Performance uninterpreted strings of bits.
Clients see the same performance
regardless of the location of objects Storage Server
and services (this is not always The module that manages the actual
achievable unless the user is willing storage of bitfiles, allocating media
to slow down local performance). extents, scheduling drives, requesting
279
volume mounts, and initiating data Mover
transfers. The Mover transmits data between two
channels. The channels can be connected
Physical Volume Repository to storage devices, host memories, or
The PVR manages physical volumes networks.
(removabledisks, magnetictapes,etc.)
and mounts them on drives, robotically Site Manager
or manually, upon request. This modules provides the administra-
tion interface to all of the other modules
of the model.
Biffile Biffile
Mover Mover
Physical
Applications Bitfile Biffile Storage Volume
Client Server Sewer
Name
All Modules
Server.
Control
Data
Volumes
The IEEE Mass Storage System Reference Model
Figure 3
The key ideas that will allow standards tiers, allowing the other modules in
based on the reference model to support the model to support a variety of
transparency are: different naming environments.
The Mover separates the data path The modularity of the Bitfile Client,
from the control path, allowing the Bitfile Server, Storage Server, and
controller-to-network path shown Physical Volume Repository allows
in Figure 1. support for different devices and
client semantics with a minimum of
The Name Server isolates the map- device- or environment-specific
ping of human-oriented names to software.
machine-oriented bitfile indenti-
28O
I would like to encourage people attending
the Goddard conference to support the IEEE
standards effort by participating in the
Storage System Standards Working Group.
For more information, contact me at:
Sam Coleman
Lawrence Livermore National Laboratory
Mail Stop L-60
P. O. Box 808
Livermore, Ca. 94550
(41 5) 422-4323
[email protected]
Until standard software systems are
available, there are steps that the storage
industry can take toward more transparent
products. The Sun Microsystems Network
File System and the CMU Andrew File
System provide a degree of transparency.
Work on these systems to improve their
security and performance, and to provide
links to hierarchical, archival systems,
will improve their transparency. I would
suggest that software vendors strive to
provide operating-system access to
archival storage systems, possibly through
mechanisms like the AT&T File System
Switch.
To learn more about all of the storage
issues that I have mentioned, I would
encourage you to attend the 11th IEEE Mass
Storage Symposium in Monterey,
California October 7-10, 1991. For
details, contact:
Bernie O'Lear
National Center for Atmospheric
Research
P. O. Box 3000
Boulder, Colorado 80307
Reference
1. Coleman, S. and Miller, S., editors, A
Reference Model for Mass Storage
Systems, IEEE Technical Committee on
Mass Storage Systems and Technology,
May, 1990.
2_1
T--
<
O3
,<
z
==
G)
0
L.
==
LI.
U.I
1.._ 0
0 _-
_-J
O_
oJ
0
09
o9
z
0
q,,,
im
283
r_
oO
_C
Z
A
T-"
0
Z
0
m
s_...
u
_3 .c
ILl
0 --"
mm
•g,mp e-
mm
U
tv..
!'- "'
..r-
,,I,,,d
"-- (_
nm ,,4rid
(_ ,,4,,d
U
18_ o • • • o o o o o o • • • • • • - ,,j