The following notable changes and features are included in this update:

• Automatic NUMA balancing disabled by default

  In this update, the auto-enabling of automatic Non-uniform memory access (NUMA) balancing has been disabled. This change addresses several issues that were encountered on systems with multiple NUMA nodes, where automatic NUMA balancing was enabled. Symptoms that were reported included high iowait times and numerous processes (like Oracle DB processes) observed in the D state, that were sitting on the wait_on_page_bit() function in the process stack.

  You can manually enable automatic NUMA balancing by setting the kernel.numa_balancing kernel parameter to 1.

  To set the kernel.numa_balancing kernel parameter at runtime, use the sysctl command, as follows:

  # sysctl -w kernel.numa_balancing=1

  To make the enabling of automatic NUMA balancing persistent across a reboot, add the kernel.numa_balancing=1 parameter to the /etc/sysctl.conf file.

  To verify whether the kernel.numa_balancing kernel parameter is set, run the following command:

  # sysctl kernel.numa_balancing

• Deferred compaction feature backported and set to madvise by default

  The deferred compaction feature for Transparent Huge Pages (THP) has been backported from the mainline kernel, and the default defrag behavior has been set to madvise. This fix resolves an issue where THP could cause application stalls in the event that a huge page allocation from THP took a long time due to memory fragmentation.

• FRWR support added in RDS

  This update adds support for MR registration through work requests in RDS, which is commonly referred to as FRWR/fastreg/FRMR. With this change, RDS can choose either FMR or FRWR as the registration method. The choice is based on the preference that is specified in the prefer_frwr module parameter, as well as the support that is offered by the underlying device.

• ibacm available in the Oracle OFED release

  Starting with the Oracle OFED 1.0.0-33 release, ibacm includes Netlink support. The ibacm service processes path record queries and also caches path records. Because ibacm provides an easy path record cache for kernel components, this implementation greatly improves performance on large fabric systems.

• libnvdimm subsystem added to kernel

  This update adds the libnvdimm kernel subsystem, which is responsible for the detection, configuration, and management of Non-Volatile Dual Inline Memory Modules (NVDIMMs). As a result, if NVDIMMs are present in the system, they are exposed through the /dev/pmem* device nodes and can be configured by using the ndctl utility.

• panic_on_rcu_stall parameter added to sysctl to debug RCU stalls on a vmcore

  The panic_on_rcu_stall interface is useful for defining the root cause of RCU stalls when using a vmcore. You can set the parameter's value to 1 to call the panic() function whenever an RCU stall is detected.

• Target Core Module in Userspace (TCMU) support for in-kernel iSCSI target

  This feature provides support for userspace pass-through, which enables userspace modules to be specified as iSCSI targets. Typically, built-in modules are used as backstores or storage engines for SCSI devices and are implemented entirely as kernel code. Backstores cover common use cases, but some target solutions can support alternate backstores and act as translators for initiators that might need to store data on these non-traditional storage systems. By supporting TCMU, a small adapter module can be released for each backstore. This functionality is similar to Filesystem in Userspace (FUSE), but at the SCSI layer rather than the file system layer.

• uio_hv_generic driver enabled

  The uio_hv-generic driver is a generic driver that can be bound to any Hyper-V VMBus device, providing there is userspace access to networking and adequate storage is available. The addition of this driver benefits both virtio and Xen drivers.