SPDK's event/application framework (
lib/event) now supports scheduling of lightweight threads. Schedulers are provided as plugins, called implementations. A default implementation is provided, but users may wish to write their own scheduler to integrate into broader code frameworks or meet their performance needs.
This feature should be considered experimental and is disabled by default. When enabled, the scheduler framework gathers data for each spdk thread and reactor and passes it to a scheduler implementation to perform one of the following actions.
spdk_threads can be moved to another reactor. Schedulers can examine the suggested cpu_mask value for each lightweight thread to see if the user has requested specific reactors, or choose a reactor using whatever algorithm they deem fit.
Reactors by default run in a mode that constantly polls for new actions for the most efficient processing. Schedulers can switch a reactor into a mode that instead waits for an event on a file descriptor. On Linux, this is implemented using epoll. This results in reduced CPU usage but may be less responsive when events occur. A reactor cannot enter this mode if any
spdk_threads are currently scheduled to it. This limitation is expected to be lifted in the future, allowing
spdk_threads to enter interrupt mode.
The frequency of CPU cores can be modified by the scheduler in response to load. Only CPU cores that match the application cpu_mask may be modified. The mechanism for controlling CPU frequency is pluggable and the default provided implementation is called
dpdk_governor, based on the
rte_power library from DPDK.
When SMT (Hyperthreading) is enabled the two logical CPU cores sharing a single physical CPU core must run at the same frequency. If one of two of such logical CPU cores is outside the application cpu_mask, the policy and frequency on that core has to be managed by the administrator.
The scheduler in use may be controlled by JSON-RPC. Please use the framework_set_scheduler RPC to switch between schedulers or change their options. Currently only dynamic scheduler supports changing its parameters.
spdk_top is a useful tool to observe the behavior of schedulers in different scenarios and workloads.
static scheduler is the default scheduler and does no dynamic scheduling. Lightweight threads are distributed round-robin among reactors, respecting their requested cpu_mask, and then they are never moved. This is equivalent to the previous behavior of the SPDK event/application framework.
dynamic scheduler is designed for power saving and reduction of CPU utilization, especially in cases where workloads show large variations over time. In SPDK thread and core workloads are measured in CPU ticks. Those values are then compared with all the ticks since the last check, which allows to calculate
busy time = busy ticks / (busy tick + idle tick) * 100 %
The thread is considered to be active, if its busy time is over the
load limit parameter.
Active threads are distributed equally among reactors, taking cpu_mask into account. All idle threads are moved to the main core. Once an idle thread becomes active, it is redistributed again. Dynamic scheduler monitors core workloads and redistributes SPDK threads on cores in a way that none of them is over
core limit. In case a core utilization surpasses this threshold, scheduler should move threads out of it until this condition no longer applies. Cores might also be in overloaded state, which indicates that moving threads out of this core will not decrease its utilization under the
core limit and the threads are unable to process all the I/O they are capable of, because they share CPU ticks with other threads. The threshold to decide if a core is overloaded is called
core busy. Note that threads residing on an overloaded core will not perform as good as other threads, because the CPU ticks intended for them are limited by other threads on the same core.
When a reactor has no scheduled
spdk_threads it is switched into interrupt mode and stops actively polling. After enough threads become active, the reactor is switched back into poll mode and threads are assigned to it again.
The main core can contain active threads only when their execution time does not exceed the sum of all idle threads. When no active threads are present on the main core, the frequency of that CPU core will decrease as the load decreases. All CPU cores corresponding to the other reactors remain at maximum frequency.
The dynamic scheduler is currently the only one that allows manual setting of its parameters.
Current values of scheduler parameters can be displayed by using framework_get_scheduler RPC.