 |
» |
|
|
|
|
|
|
|
This article describes
the results of tests performed by Resilient Systems at Hewlett-Packard’s
Littleton, Massachusetts laboratory in the spring of 2003, using hardware and
test suites provided by OpenVMS Engineering. The tests were performed on a
16-way GS1280 AlphaServer running multiple instances of CHARON-VAX/AXP Plus.
The test suite was the
same one that OpenVMS Engineering used in previous decades to test new VAX
hardware designs. To ensure proper execution of the VAX instruction set, the
tests verify conformance of the new hardware to expected test results. The
comprehensive suite exercises nearly every VAX instruction, including all
three-operand VAX instructions as well as single, double, and floating-point
calculation speeds. For some instructions, the tests revealed that the
CHARON-VAX/AXP Plus emulator was more than ten times faster than any real VAX
processor.
In addition to this
suite, Resilient Systems used the VUPs Calculator utility to test individual
CPU performance in processing a mix of fixed and floating point instructions.
To test scalability of CHARON-VAX/AXP Plus on multiprocessor configurations,
Resilient used standard Dhrystone tests because the results are much more
granular than those produced by the VUPs Calculator. Resilient Systems first
determined peak Dhrystones on a single CPU and then ran simultaneous Dhrystone
tests on multiple instances of CHARON-VAX/AXP Plus in multiple n-way
GS1280 configurations, with up to 18 instances on a 16-way GS1280. Resilient
Systems repeated these tests over three days, and then calculated the average
Dhrystone performance. Resilient Systems then used the standard formula to
convert Dhrystones to VUPs, and produced detailed graphs to show the results
(the graphs are shown later in this article in Figure 4 and Figure
5).
The complete results
from the OpenVMS Engineering test suite; the CPU, disk, and network performance
tests; and the CHARON-VAX/AXP Plus
scalability tests are available from Resilient Systems.
The next two sections
describe the benefits of (1) the latest release of HP AlphaServer products (and
the GS1280 model in particular) and (2) the CHARON-VAX/AXP Plus emulator, respectively. The third section to follow provides
details about the cumulative performance benefits of these two products
combined.
|
|
High-Performance Chip Technology and Scaling Advancements |
|
|
|
All elements required
for symmetric multiprocessing now reside on a single chip. In addition to an
on-chip L2 cache, two on-chip memory controllers provide exceptional memory
bandwidth. In an industry-first achievement, an on-chip router connects
AlphaServer processors directly to one another. This “switchless” mesh design
results in a very high interconnect bandwidth of up to 64 CPUs. SPEC_rate 2000
tests on a 32-way GS1280 system proved that the GS1280 can achieve nearly 100%
linear scalability. In other words, as the number of processors used on the
GS1280 increases, the performance increases at a constant rate.
The I/O performance
and scaling improvements of the GS1280 over the GS320 are equally impressive.
The GS1280 provides flexibility in configuring I/O, from one I/O chip per
system to one I/O chip per processor.
The result is a platform with linear scaling in I/O, yielding eight
times the I/O bandwidth of the GS320. Moreover, the GS1280 Lego™ block design
of hot-swappable components results in a robust platform with 15% to 30%
improvement in Mean Time Between Failure (MTBF) over the previous generation of
AlphaServers. Available in multiple processor (n-way) configurations, the enterprise-scale AlphaServer GS1280,
along with the departmental and workgroup ES80 and ES47 models, provide
significant performance and reliability improvements over the earlier GS320 and
ES45 models.
|
Partitioning Enables Support of Mixed-Architecture Clusters in a Single AlphaServer Box |
|
|
|
With the new ES47,
ES80, and GS1280 AlphaServer products, HP introduced support for hardware
partitions. Hardware partitions permit multiple instances of the OpenVMS
operating system to run concurrently in physically separate parts of the
computer. Such a configuration facilitates the dedication of partitions to
specific applications, with the ability to tune and secure each partition to
the specific demands of its application set. By effecting the partition of the
system into multiple independent Alpha processors, this new feature facilitates
the deployment and execution of multiple instances of CHARON-VAX/AXP Plus. CHARON-VAX/AXP Plus can be run as an application on one
or more of the CPUs in a processor partition, or across multiple CPUs in
multiple partitions.
This allows
construction of a variety of mixed VAX and Alpha configurations, all within a
single system cabinet. For example, by using the 8-way AlphaServer depicted in Figure 1, we can partition it into three separate Alpha
systems, forming a consolidated mixed-architecture configuration as shown in
Figure 2. The Alpha system constructed from the five CPUs (shown in yellow)
could be configured to run multiple CHARON-VAX/AXP Plus instances.
Figure 1: Multi-CPU Alpha Processor
The other two
AlphaServer partitions could be added to create the 6-node, mixed-architecture
cluster, shown in Figure 2. In this configuration, three of the eight available
CPUs would run as actual Alpha nodes — one single CPU node (blue) and one
dual-CPU multi-processor node (pink). The remaining five CPUs (yellow) would
run four instances of CHARON-VAX/AXP Plus
as four VAX nodes, with the fifth CPU (labeled in the figure as the “Alpha
Management CPU”) fielding user interrupts and managing disk and network I/O as
described in the next section.
Figure 2: Mixed VAX and Alpha Cluster Example
|
|
|
|
|
The new VAX-on-Alpha
emulator from Software Resources International takes full advantage of the
revolutionary improvements of HP’s EV7-based AlphaServer systems. Software
Resources International specializes in migrating operating systems and
applications to modern platforms (for example, migrating OpenVMS Alpha systems
to OpenVMS I64 systems based on the Intel Itanium architecture) and developing
hardware emulators for PDP and VAX processors. The emulators are mathematical
models of the hardware architecture; written in C, they run as ordinary
applications on modern platforms.
Figure 3 illustrates how the Software Resources International
CHARON-VAX/AXP Plus emulator running
on an OpenVMS Alpha system replaces the OpenVMS VAX hardware, providing the
same operating system functionality and application support. The OpenVMS VAX
software, layered software, and user applications are installed onto the CHARON-VAX/AXP Plus emulator which is running on the AlphaServer, which in turn is
running its own copy of OpenVMS. With CHARON-VAX/AXP Plus, no conversion of code is needed. Simply use BACKUP/IMAGE to
transfer existing OpenVMS system and application binaries to the CHARON-VAX/AXP
emulator running on the OpenVMS Alpha system, as if you were simply moving from
one VAX model to another.
Figure 3: CHARON-VAX/AXP Plus on an HP AlphaServer Easily Replaces VAX Hardware
The software
architecture of the CHARON-VAX/AXP Plus
emulator consists of two threads — one thread to execute the emulator and a
second thread to field interrupts, run the scheduler, manage resources, handle
I/O to storage devices, and manage network I/O. While you can run both threads on the same processor, for optimum
performance the emulator thread should have 100% of a CPU available to it. The
second thread, automatically assigned to a separate CPU when one is available,
requires a fraction of the compute power available to it.
CHARON-VAX/AXP Plus is the second-generation of
Software Resources International’s VAX hardware emulator for Alpha. The first
emulator modeled a MicroVAX 3600. The new emulator provides the functionality
of a VAX 3100 Model 98 hardware system, complete with up to 512 MB memory, dual
SCSI storage buses, and a 10/100 Mbps Ethernet network.
Combined with the
scalability and reliability of the ES47, ES80, and GS1280 AlphaServer products,
the sophisticated instruction preprocessing now provided by the CHARON-VAX/AXP Plus emulator has significantly
increased the viability of preserving business-critical VAX applications by
means of VAX emulation. As a result, one or many low to mid-range MicroVAX
processors can be replaced by entry-level ES47s. Testing has now shown that the
CHARON-VAX/AXP Plus emulator running
on an n-way GS1280 can replace one or
more high-end VAX processors, such as VAX 77xx’s
or VAX 78xx’s.
|
|
|
|
|
|
Specifically, the
testing conducted by Resilient Systems at HP’s Littleton, Massachusetts
laboratory proved that a 16-way GS1280 running CHARON-VAX/AXP Plus delivers the equivalent of a VAX
3198 or VAX 7610 (over 36 VAX units of performance (VUPs)) on each CPU
of the AlphaServer system.
Even more impressive,
the remarkably efficient CHARON-VAX kernel (0.5 MB) achieved the same
scalability as the underlying GS1280 hardware when running multiple instances
of CHARON-VAX/AXP Plus. As the graph shows
in Figure
4, the compute power (measured in VUPs) obtained by
running multiple instances of the VAX emulator scaled nearly linearly. In other
words, each additional emulator instance adds nearly the same amount of compute
power to the cumulative stacked bar graph even though the instances are all
competing for resources from the same Alpha management CPU.
Figure 4: Compute Power of Multiple Instances of CHARON-VAX/AXP PlusScales Nearly Linearly
The remarkable synergy
between the hardware architecture of the HP GS1280 AlphaServer and the software
architecture of CHARON-VAX/AXP Plus
produced an optimum configuration of 15 instances of the emulator on a 16-way
AlphaServer, with the 16th CPU managing resources for the other 15.
Specifically, tests
proved each instance of CHARON-VAX/AXP Plus
delivers an average of 32 VUPs on an AlphaServer with 16 CPUs, each CPU
independently running an instance of the emulator. As the graph demonstrates in
Figure
5, performance remained above 30 VUPs per CPU except
when the number of CHARON-VAX executables exceeded the physical number of CPUs.
Figure 5: Incremental VUPs per Instance of the CHARON-VAX/AXP Plus Emulator
|
|
|
|
|
|
The results of these
tests clearly indicate that multiple individual VAX servers or VAXstations
could easily be consolidated on the same GS1280 AlphaServer host. Server consolidation offers many
benefits to VAX sites with multiple systems, including reduced footprint and
power consumption, and greatly reduced hardware maintenance costs. In addition,
the superb reliability and MTBF of the GS1280 AlphaServer reduces staffing
requirements and dependence on increasingly scarce VAX/VMS system engineers,
while also reducing the risk of
business disruption due to malfunction of aging hardware.
Similarly, single platform clusters can now be
created — an entire cluster of existing VAX processors could be recreated as
multiple cluster members of the same cluster, all residing and managed on a
single GS1280 host.
Alternatively, the
configuration could be aggregated and then spread over redundant GS1280 systems
to attain the highest possible availability through the independence of
separate hardware systems. The benefits would include all the benefits of
server consolidation described above, plus the failover capability inherent in
OpenVMS Clusters.
Note that in either
scenario — server consolidation or single platform clusters — you must
carefully calculate your VUP performance needs to ensure that the number of
CHARON-VAX/AXP Plus instances on the
GS1280 stays within the recommended limits. The chart in Figure 6 illustrates the cumulative capacity (in VUPs) of the
various AlphaServer platforms, with the maximum capacity being provided by the
GS1280 (16 instances producing approximately 460 VUPs).
Figure 6: Cumulative Capacity of Various AlphaServer Platforms Running Multiple Instances of CHARON-VAX/AXP Plus
|
|
|
|
|
|
As more evidence of
the synergy between the GS1280 hardware and the CHARON-VAX/AXP Plus software, tests show that the
bandwidth available to the emulator is nearly identical to what is physically
attached to the AlphaServer host. Repeated testing showed that native Alpha
disk transfers achieved 4.47 MB/sec when accessing a local SCSI disk versus
4.45 MB/sec for the CHARON-VAX/AXP Plus
emulator when accessing the same physical disk. In other words, emulator
overhead is less than 1% for tasks such as disk-to-disk file copy operations or
OpenVMS backup transfers.
The tests prove that
customers are now able to assimilate high performance storage subsystems, such
as Fibre Channel, into a legacy OpenVMS VAX configuration. The CHARON-VAX/AXP Plus software increases storage capacity
by transforming VAX physical disks into disk image files on the replacement
platform. Now, with the HP AlphaServer’s support for robust storage
technologies, the CHARON-VAX/AXP Plus
software enables critical VAX applications to take advantage of both storage and
I/O throughput capacities that were unimaginable in the heyday of the VAX
processor.
The integrated
Ethernet adaptor emulator provided with the CHARON-VAX/AXP Plus product is pivotal to integrating an instance of the emulator
with other DECnet nodes and cluster members, or by means of IP, with corporate
LANs and WANs. This channel also provides user connectivity through Telnet and
third-party terminal emulators. Thus, it is a key component of a VAX
replacement configuration. When the adaptor emulators were set in tests to
match the 10baseT adapter of a VAX system, Resilient Systems observed data
rates through the network device at over 1.8MB/sec for sustained data transfer,
and near the full 10MB/sec possible for message transfer. You can use 100 Mbps
Ethernet adapters with the current version of CHARON-VAX/AXP Plus, but these were not tested.
Operation at 100 Mbps requires an Alpha SMP host with a CPU frequency of at
least 1 GHz. Network throughput can be tuned individually for specific protocol
classes (for example, DECnet, TCP/IP,
or OpenVMS Cluster communication).
|
|
|
