One way to improve the performance of an OSA-Express3 or later
feature in QDIO mode for processing workloads with demanding latency
requirements is to configure the OSA-Express feature to operate in
optimized latency mode. Optimized latency mode optimizes interrupt
processing for both inbound and outbound data, which decreases latency
and can provide significant increases in throughput, particularly
for high volume, interactive, non-streaming workloads.
To configure an OSA-Express3 or later feature to operate in optimized
latency mode, use the INTERFACE statement with the OLM parameter.
Because optimized latency mode affects both inbound and outbound interrupts,
it supersedes other inbound performance settings that the INBPERF
parameter sets. For more information about optimized latency mode
and the OLM and INBPERF parameters on the INTERFACE statement for IPAQENET and IPAQENET6, see z/OS Communications Server: IP Configuration
Reference.
Because of the operating characteristics of optimized latency mode,
two other configuration changes might be required.
- For outbound traffic to gain the benefit of optimized latency
mode, direct traffic to priority queues 1, 2, or 3 by using the WLMPRIORITYQ
parameter on the GLOBALCONFIG statement, or by using Policy Agent
and configuring a policy with the SetSubnetPrioTosMask statement.
Although an OSA-Express feature supports multiple outbound write
priority queues, outbound optimized latency mode occurs only for traffic
on priority queue 1 (priority level 1). The TCP/IP stack combines
all the traffic that is directed to priority queues 1, 2, and 3 into
priority queue 1 for any OSA-Express3 or later feature that is operating
in optimized latency mode.
For more information about directing
traffic to outbound OSA-Express priority queues by using the WLMPRIORITYQ
parameter on the GLOBALCONFIG statement or using the SetSubnetPrioTosMask statement, see z/OS Communications Server: IP Configuration
Reference.
Guideline: Configure
the WLMPRIORITYQ parameter with no subparameters, which assigns a
default mapping of service class importance levels to OSA-Express
outbound priority queues. This default mapping directs traffic that
is assigned to the higher priority service class importance levels
1–4 to queues that operate in optimized latency mode, and enables
the appropriate types of traffic to benefit from optimized latency
mode.
Result: If neither the
WLMPRIORITYQ or SetSubnetPrioTosMask statements are specified, any
packet that has a type of service (ToS) byte with the first 3 bits
being 000 or 001 is directed to queue 4 and does not benefit from
optimized latency mode.
- To achieve optimal latency for one or more network interfaces
that are operating in optimized latency mode, limit the number of
network interfaces that can concurrently share an OSA-Express3 or
later feature.
When at least one network interface is operating
in optimized latency mode, ensure that no more than four concurrent
network interfaces are sharing an OSA-Express3 or later port, and
no more than eight concurrent network interfaces are sharing an OSA-Express3
or later channel path identifier (CHPID). The following configurations
can result in multiple users that are sharing an OSA-Express3 or later
feature.
- Multiple LPARs sharing the OSA-Express3 or later feature
- Multiple stacks on the same LPAR sharing the OSA-Express3 or later
feature
- Multiple VLAN interfaces to the OSA-Express3 or later feature
- Both an IPv4 and an IPv6 active interface to the OSA-Express3
or later feature
- A TCP/IP stack that is enabling the OSA-Express network traffic
analyzer (OSAENTA) for the OSA-Express3 or later feature
Optimized latency mode is intended for high volume, interactive
workloads. Although optimized latency mode can compensate for some
mixing of workloads, an excessive amount of high volume, streaming
workloads, such as bulk data or file transfer, can result in higher
processor consumption.
Guideline: When multipath routing by
using the PERPACKET option is enabled, do not configure a multipath
group that contains an OSA-Express3 or later feature configured with
optimized latency mode and any other type of device.
Restrictions: - Optimized latency mode is limited to OSA-Express3 or later Ethernet
features in QDIO mode that are running with an IBM® System z10® or later server. For more information, see the 2097DEVICE
Preventive Service Planning (PSP) bucket.
- Traffic that is either inbound over or being forwarded to an OSA-Express3
or later feature configured to operate in optimized latency mode is
not eligible for the accelerated routing that HiperSockets™ Accelerator and QDIO Accelerator
provide.
- For an OSA-Express3 or later interface configured to operate in
optimized latency mode, the stack ignores the configured INBPERF setting
and uses the value DYNAMIC.