Guidelines for OpenVMS Cluster Configurations
10.7.8 Extended LAN Configuration Guidelines
You can use bridges and switches between LAN segments to form an
extended LAN. This can increase availability, distance, and aggregate
bandwidth as compared with a single LAN. However, an extended LAN can
increase delay and can reduce bandwidth on some paths. Factors such as
packet loss, queuing delays, and packet size can also affect network
performance. Table 10-3 provides guidelines for ensuring adequate
LAN performance when dealing with such factors.
Table 10-3 Extended LAN Configuration Guidelines
| Factor |
Guidelines |
|
Propagation delay
|
The amount of time it takes a packet to traverse the LAN depends on the
distance it travels and the number of times it is relayed from one link
to another through a switch or bridge. If responsiveness is critical,
then you must control these factors.
When an FDDI is used for OpenVMS Cluster communications, the ring
latency when the FDDI ring is idle should not exceed 400 ms. FDDI
packets travel at 5.085 microseconds/km and each station causes an
approximate 1-ms delay between receiving and transmitting. You can
calculate FDDI latency by using the following algorithm:
Latency = (distance in km) * (5.085 ms/km) + (number of stations) *
(1 ms/station)
For high-performance applications, limit the number of switches between
nodes to two. For situations in which high performance is not required,
you can use up to seven switches or bridges between nodes.
|
|
Queuing delay
|
Queuing occurs when the instantaneous arrival rate at switches or
bridges and host adapters exceeds the service rate. You can control
queuing by:
- Reducing the number of switches or bridges between nodes that
communicate frequently.
- Using only high-performance switches or bridges and adapters.
- Reducing traffic bursts in the LAN. In some cases, for example, you
can tune applications by combining small I/Os so that a single packet
is produced rather than a burst of small ones.
- Reducing LAN segment and host processor utilization levels by using
faster processors and faster LANs, and by using switches or bridges for
traffic isolation.
|
|
Packet loss
|
Packets that are not delivered by the LAN require retransmission, which
wastes network resources, increases delay, and reduces bandwidth.
Bridges and adapters discard packets when they become congested. You
can reduce packet loss by controlling queuing, as previously described.
Packets are also discarded when they become damaged in transit. You
can control this problem by observing LAN hardware configuration rules,
removing sources of electrical interference, and ensuring that all
hardware is operating correctly.
Packet loss can also be reduced by using VMS Version 5.5--2 or later,
which has PEDRIVER congestion control.
The retransmission timeout rate, which is a symptom of packet loss,
must be less than 1 timeout in 1000 transmissions for OpenVMS Cluster
traffic from one node to another. LAN paths that are used for
high-performance applications should have a significantly lower rate.
Monitor the occurrence of retransmission timeouts in the OpenVMS
Cluster.
Reference: For information about monitoring the
occurrence of retransmission timeouts, see HP OpenVMS Cluster Systems.
|
|
Switch or bridge recovery delay
|
Choose switches or bridges with fast self-test time and adjust them for
fast automatic reconfiguration. This includes adjusting spanning tree
parameters to match network requirements.
Reference: Refer to HP OpenVMS Cluster Systems for more information
about LAN bridge failover.
|
|
Bandwidth
|
All LAN paths used for OpenVMS Cluster communication must operate with
a nominal bandwidth of at least 10 Mb/s. The average LAN segment
utilization should not exceed 60% for any 10-second interval.
For FDDI configurations, use FDDI exclusively on the communication
paths that have the highest performance requirements. Do not put an
Ethernet LAN segment between two FDDI segments. FDDI bandwidth is
significantly greater, and the Ethernet LAN will become a bottleneck.
This strategy is especially ineffective if a server on one FDDI must
serve clients on another FDDI with an Ethernet LAN between them. A more
appropriate strategy is to put a server on an FDDI and put clients on
an Ethernet LAN, as Figure 10-21 shows.
For Gigabit Ethernet configurations, enable jumbo frames where
possible.
|
|
Traffic isolation
|
Use switches or bridges to isolate and localize the traffic between
nodes that communicate with each other frequently. For example, use
switches or bridges to separate the OpenVMS Cluster from the rest of
the LAN and to separate nodes within an OpenVMS Cluster that
communicate frequently from the rest of the OpenVMS Cluster.
Provide independent paths through the LAN between critical systems
that have multiple adapters.
|
|
Packet size
|
You can adjust the NISCS_MAX_PKTSZ system parameter to use the full
FDDI packet size. Ensure that the LAN path supports a data field of at
least 4474 bytes end to end.
Some failures cause traffic to switch from an LAN path that
supports a large packet size to a path that supports only smaller
packets. It is possible to implement automatic detection and recovery
from these kinds of failures. This capability requires that the ELAN
set the value of the priority field in the FDDI frame-control byte to
zero when the packet is delivered on the destination FDDI link.
Ethernet-to-FDDI bridges that conform to the IEEE 802.1 bridge
specification provide this capability.
|
10.7.9 System Parameters for OpenVMS Clusters
In an OpenVMS Cluster with satellites and servers, specific system
parameters can help you manage your OpenVMS Cluster more efficiently.
Table 10-4 gives suggested values for these system parameters.
Table 10-4 OpenVMS Cluster System Parameters
| System Parameter |
Value for
Satellites |
Value for
Servers |
|
LOCKDIRWT
|
0
|
1-4. The setting of LOCKDIRWT influences a node's willingness to serve
as a resource directory node and also may be used to determine
mastership of resource trees. In general, a setting greater than 1 is
determined after careful examination of a cluster node's specific
workload and application mix and is beyond the scope of this document.
|
|
SHADOW_MAX_COPY
|
0
|
4, where a significantly higher setting may be appropriate for your
environment
|
|
MSCP_LOAD
|
0
|
1
|
|
NPAGEDYN
|
Higher than for standalone node
|
Higher than for satellite node
|
|
PAGEDYN
|
Higher than for standalone node
|
Higher than for satellite node
|
|
VOTES
|
0
|
1
|
|
EXPECTED_VOTES
|
Sum of OpenVMS Cluster votes
|
Sum of OpenVMS Cluster votes
|
|
RECNXINTERVL
1
|
Equal on all nodes
|
Equal on all nodes
|
1Correlate with bridge timers and LAN utilization.
Reference: For more information about these
parameters, see HP OpenVMS Cluster Systems and HP Volume Shadowing for OpenVMS.
10.8 Scaling for I/Os
The ability to scale I/Os is an important factor in the growth of your
OpenVMS Cluster. Adding more components to your OpenVMS Cluster
requires high I/O throughput so that additional components do not
create bottlenecks and decrease the performance of the entire OpenVMS
Cluster. Many factors can affect I/O throughput:
- Direct access or MSCP served access to storage
- Settings of the MSCP_BUFFER and MSCP_CREDITS system parameters
- File system technologies, such as Files-11
- Disk technologies, such as magnetic disks, solid-state disks, and
DECram
- Read/write ratio
- I/O size
- Caches and cache "hit" rate
- "Hot file" management
- RAID striping and host-based striping
- Volume shadowing
These factors can affect I/O scalability either singly or in
combination. The following sections explain these factors and suggest
ways to maximize I/O throughput and scalability without having to
change in your application.
Additional factors that affect I/O throughput are types of
interconnects and types of storage subsystems.
Reference: For more information about interconnects,
see Chapter 4. For more information about types of storage
subsystems, see Chapter 5. For more information about MSCP_BUFFER
and MSCP_CREDITS, see HP OpenVMS Cluster Systems.
10.8.1 MSCP Served Access to Storage
MSCP server capability provides a major benefit to OpenVMS Clusters: it
enables communication between nodes and storage that are not directly
connected to each other. However, MSCP served I/O does incur overhead.
Figure 10-24 is a simplification of how packets require extra handling
by the serving system.
Figure 10-24 Comparison of Direct and MSCP Served
Access
In Figure 10-24, an MSCP served packet requires an extra
"stop" at another system before reaching its destination.
When the MSCP served packet reaches the system associated with the
target storage, the packet is handled as if for direct access.
In an OpenVMS Cluster that requires a large amount of MSCP serving, I/O
performance is not as efficient and scalability is decreased. The total
I/O throughput is approximately 20% less when I/O is MSCP served than
when it has direct access. Design your configuration so that a few
large nodes are serving many satellites rather than satellites serving
their local storage to the entire OpenVMS Cluster.
10.8.2 Disk Technologies
In recent years, the ability of CPUs to process information has far
outstripped the ability of I/O subsystems to feed processors with data.
The result is an increasing percentage of processor time spent waiting
for I/O operations to complete.
Solid-state disks (SSDs), DECram, and RAID level 0 bridge this gap
between processing speed and magnetic-disk access speed. Performance of
magnetic disks is limited by seek and rotational latencies, while SSDs
and DECram use memory, which provides nearly instant access.
RAID level 0 is the technique of spreading (or "striping") a
single file across several disk volumes. The objective is to reduce or
eliminate a bottleneck at a single disk by partitioning heavily
accessed files into stripe sets and storing them on multiple devices.
This technique increases parallelism across many disks for a single I/O.
Table 10-5 summarizes disk technologies and their features.
Table 10-5 Disk Technology Summary
| Disk Technology |
Characteristics |
|
Magnetic disk
|
Slowest access time.
Inexpensive.
Available on multiple interconnects.
|
|
Solid-state disk
|
Fastest access of any I/O subsystem device.
Highest throughput for write-intensive files.
Available on multiple interconnects.
|
|
DECram
|
Highest throughput for small to medium I/O requests.
Volatile storage; appropriate for temporary read-only files.
Available on any Alpha or VAX system.
|
|
RAID level 0
|
Available on HSD, HSJ, and HSG controllers.
|
Note: Shared, direct access to a solid-state disk or
to DECram is the fastest alternative for scaling I/Os.
10.8.3 Read/Write Ratio
The read/write ratio of your applications is a key factor in scaling
I/O to shadow sets. MSCP writes to a shadow set are duplicated on the
interconnect.
Therefore, an application that has 100% (100/0) read activity may
benefit from volume shadowing because shadowing causes multiple paths
to be used for the I/O activity. An application with a 50/50 ratio will
cause more interconnect utilization because write activity requires
that an I/O be sent to each shadow member. Delays may be caused by the
time required to complete the slowest I/O.
To determine I/O read/write ratios, use the DCL command MONITOR IO.
10.8.4 I/O Size
Each I/O packet incurs processor and memory overhead, so grouping I/Os
together in one packet decreases overhead for all I/O activity. You can
achieve higher throughput if your application is designed to use bigger
packets. Smaller packets incur greater overhead.
10.8.5 Caches
Caching is the technique of storing recently or frequently used data in
an area where it can be accessed more easily---in memory, in a
controller, or in a disk. Caching complements solid-state disks,
DECram, and RAID. Applications automatically benefit from the
advantages of caching without any special coding. Caching reduces
current and potential I/O bottlenecks within OpenVMS Cluster systems by
reducing the number of I/Os between components.
Table 10-6 describes the three types of caching.
Table 10-6 Types of Caching
| Caching Type |
Description |
|
Host based
|
Cache that is resident in the host system's memory and services I/Os
from the host.
|
|
Controller based
|
Cache that is resident in the storage controller and services data for
all hosts.
|
|
Disk
|
Cache that is resident in a disk.
|
Host-based disk caching provides different benefits from
controller-based and disk-based caching. In host-based disk caching,
the cache itself is not shareable among nodes. Controller-based and
disk-based caching are shareable because they are located in the
controller or disk, either of which is shareable.
10.8.6 Managing "Hot" Files
A "hot" file is a file in your system on which the most
activity occurs. Hot files exist because, in many environments,
approximately 80% of all I/O goes to 20% of data. This means that, of
equal regions on a disk drive, 80% of the data being transferred goes
to one place on a disk, as shown in Figure 10-25.
Figure 10-25 Hot-File Distribution
To increase the scalability of I/Os, focus on hot files, which can
become a bottleneck if you do not manage them well. The activity in
this area is expressed in I/Os, megabytes transferred, and queue depth.
RAID level 0 balances hot-file activity by spreading a single file over
multiple disks. This reduces the performance impact of hot files.
Use the following DCL commands to analyze hot-file activity:
- MONITOR IO command---Monitors hot disks.
- MONITOR MSCP command---Monitors MSCP servers.
The MONITOR IO and the MONITOR MSCP commands enable you to find out
which disk and which server are hot.
10.8.7 Volume Shadowing
The Volume Shadowing for OpenVMS product ensures that data is available
to applications and end users by duplicating data on multiple disks.
Although volume shadowing provides data redundancy and high
availability, it can affect OpenVMS Cluster I/O on two levels:
| Factor |
Effect |
|
Geographic distance
|
Host-based volume shadowing enables shadowing of any devices in an
OpenVMS Cluster system, including those served by MSCP servers. This
ability can allow great distances along with MSCP overhead. For
example, OpenVMS Cluster systems using FDDI can be located up to 25
miles (40 kilometers) apart. Using Fibre Channel, they can be located
up to 62 miles (100 kilometers) apart. Both the distance and the MSCP
involvement can slow I/O throughput.
|
|
Read/write ratio
|
Because shadowing writes data to multiple volumes, applications that
are write intensive may experience reduced throughput. In contrast,
read-intensive applications may experience increased throughput because
the shadowing software selects one disk member from which it can
retrieve the data most efficiently.
|
|
