2. Driver for the Intel® Dynamic Load Balancer (DLB)

The DPDK DLB poll mode driver supports the Intel® Dynamic Load Balancer, hardware versions 2.0 and 2.5.

2.1. Prerequisites

Follow the DPDK Getting Started Guide for Linux to setup the basic DPDK environment.

2.2. Configuration

The DLB PF PMD is a user-space PMD that uses VFIO to gain direct device access. To use this operation mode, the PCIe PF device must be bound to a DPDK-compatible VFIO driver, such as vfio-pci.

2.3. Eventdev API Notes

The DLB PMD provides the functions of a DPDK event device; specifically, it supports atomic, ordered, and parallel scheduling events from queues to ports. However, the DLB hardware is not a perfect match to the eventdev API. Some DLB features are abstracted by the PMD such as directed ports.

In general the DLB PMD is designed for ease-of-use and does not require a detailed understanding of the hardware, but these details are important when writing high-performance code. This section describes the places where the eventdev API and DLB misalign.

2.3.1. Scheduling Domain Configuration

DLB supports 32 scheduling domains. When one is configured, it allocates load-balanced and directed queues, ports, credits, and other hardware resources. Some resource allocations are user-controlled – the number of queues, for example – and others, like credit pools (one directed and one load-balanced pool per scheduling domain), are not.

The DLB is a closed system eventdev, and as such the nb_events_limit device setup argument and the per-port new_event_threshold argument apply as defined in the eventdev header file. The limit is applied to all enqueues, regardless of whether it will consume a directed or load-balanced credit.

2.3.2. Load-Balanced Queues

A load-balanced queue can support atomic and ordered scheduling, or atomic and unordered scheduling, but not atomic and unordered and ordered scheduling. A queue’s scheduling types are controlled by the event queue configuration.

If the user sets the RTE_EVENT_QUEUE_CFG_ALL_TYPES flag, the nb_atomic_order_sequences determines the supported scheduling types. With non-zero nb_atomic_order_sequences, the queue is configured for atomic and ordered scheduling. In this case, RTE_SCHED_TYPE_PARALLEL scheduling is supported by scheduling those events as ordered events. Note that when the event is dequeued, its sched_type will be RTE_SCHED_TYPE_ORDERED. Else if nb_atomic_order_sequences is zero, the queue is configured for atomic and unordered scheduling. In this case, RTE_SCHED_TYPE_ORDERED is unsupported.

If the RTE_EVENT_QUEUE_CFG_ALL_TYPES flag is not set, schedule_type dictates the queue’s scheduling type.

The nb_atomic_order_sequences queue configuration field sets the ordered queue’s reorder buffer size. DLB has 2 groups of ordered queues, where each group is configured to contain either 1 queue with 1024 reorder entries, 2 queues with 512 reorder entries, and so on down to 32 queues with 32 entries.

When a load-balanced queue is created, the PMD will configure a new sequence number group on-demand if num_sequence_numbers does not match a pre-existing group with available reorder buffer entries. If all sequence number groups are in use, no new group will be created and queue configuration will fail. (Note that when the PMD is used with a virtual DLB device, it cannot change the sequence number configuration.)

The queue’s nb_atomic_flows parameter is ignored by the DLB PMD, because the DLB does not limit the number of flows a queue can track. In the DLB, all load-balanced queues can use the full 16-bit flow ID range.

2.3.3. Load-balanced and Directed Ports

DLB ports come in two flavors: load-balanced and directed. The eventdev API does not have the same concept, but it has a similar one: ports and queues that are singly-linked (i.e. linked to a single queue or port, respectively).

The rte_event_dev_info_get() function reports the number of available event ports and queues (among other things). For the DLB PMD, max_event_ports and max_event_queues report the number of available load-balanced ports and queues, and max_single_link_event_port_queue_pairs reports the number of available directed ports and queues.

When a scheduling domain is created in rte_event_dev_configure(), the user specifies nb_event_ports and nb_single_link_event_port_queues, which control the total number of ports (load-balanced and directed) and the number of directed ports. Hence, the number of requested load-balanced ports is nb_event_ports - nb_single_link_event_ports. The nb_event_queues field specifies the total number of queues (load-balanced and directed). The number of directed queues comes from nb_single_link_event_port_queues, since directed ports and queues come in pairs.

When a port is setup, the RTE_EVENT_PORT_CFG_SINGLE_LINK flag determines whether it should be configured as a directed (the flag is set) or a load-balanced (the flag is unset) port. Similarly, the RTE_EVENT_QUEUE_CFG_SINGLE_LINK queue configuration flag controls whether it is a directed or load-balanced queue.

Load-balanced ports can only be linked to load-balanced queues, and directed ports can only be linked to directed queues. Furthermore, directed ports can only be linked to a single directed queue (and vice versa), and that link cannot change after the eventdev is started.

The eventdev API does not have a directed scheduling type. To support directed traffic, the DLB PMD detects when an event is being sent to a directed queue and overrides its scheduling type. Note that the originally selected scheduling type (atomic, ordered, or parallel) is not preserved, and an event’s sched_type will be set to RTE_SCHED_TYPE_ATOMIC when it is dequeued from a directed port.

Finally, even though all 3 event types are supported on the same QID by converting unordered events to ordered, such use should be discouraged as much as possible, since mixing types on the same queue uses valuable reorder resources, and orders events which do not require ordering.

2.3.4. Flow ID

The flow ID field is preserved in the event when it is scheduled in the DLB.

2.3.5. Hardware Credits

DLB uses a hardware credit scheme to prevent software from overflowing hardware event storage, with each unit of storage represented by a credit. A port spends a credit to enqueue an event, and hardware refills the ports with credits as the events are scheduled to ports. Refills come from credit pools.

For DLB v2.5, there is a single credit pool used for both load balanced and directed traffic.

For DLB v2.0, each port is a member of both a load-balanced credit pool and a directed credit pool. The load-balanced credits are used to enqueue to load-balanced queues, and directed credits are used for directed queues. These pools’ sizes are controlled by the nb_events_limit field in struct rte_event_dev_config. The load-balanced pool is sized to contain nb_events_limit credits, and the directed pool is sized to contain nb_events_limit/4 credits. The directed pool size can be overridden with the num_dir_credits devargs argument, like so:

--allow ea:00.0,num_dir_credits=<value>

This can be used if the default allocation is too low or too high for the specific application needs. The PMD also supports a devarg that limits the max_num_events reported by rte_event_dev_info_get():

--allow ea:00.0,max_num_events=<value>

By default, max_num_events is reported as the total available load-balanced credits. If multiple DLB-based applications are being used, it may be desirable to control how many load-balanced credits each application uses, particularly when application(s) are written to configure nb_events_limit equal to the reported max_num_events.

Each port is a member of both credit pools. A port’s credit allocation is defined by its low watermark, high watermark, and refill quanta. These three parameters are calculated by the DLB PMD like so:

  • The load-balanced high watermark is set to the port’s enqueue_depth. The directed high watermark is set to the minimum of the enqueue_depth and the directed pool size divided by the total number of ports.
  • The refill quanta is set to half the high watermark.
  • The low watermark is set to the minimum of 16 and the refill quanta.

When the eventdev is started, each port is pre-allocated a high watermark’s worth of credits. For example, if an eventdev contains four ports with enqueue depths of 32 and a load-balanced credit pool size of 4096, each port will start with 32 load-balanced credits, and there will be 3968 credits available to replenish the ports. Thus, a single port is not capable of enqueueing up to the nb_events_limit (without any events being dequeued), since the other ports are retaining their initial credit allocation; in short, all ports must enqueue in order to reach the limit.

If a port attempts to enqueue and has no credits available, the enqueue operation will fail and the application must retry the enqueue. Credits are replenished asynchronously by the DLB hardware.

2.3.6. Software Credits

The DLB is a “closed system” event dev, and the DLB PMD layers a software credit scheme on top of the hardware credit scheme in order to comply with the per-port backpressure described in the eventdev API.

The DLB’s hardware scheme is local to a queue/pipeline stage: a port spends a credit when it enqueues to a queue, and credits are later replenished after the events are dequeued and released.

In the software credit scheme, a credit is consumed when a new (.op = RTE_EVENT_OP_NEW) event is injected into the system, and the credit is replenished when the event is released from the system (either explicitly with RTE_EVENT_OP_RELEASE or implicitly in dequeue_burst()).

In this model, an event is “in the system” from its first enqueue into eventdev until it is last dequeued. If the event goes through multiple event queues, it is still considered “in the system” while a worker thread is processing it.

A port will fail to enqueue if the number of events in the system exceeds its new_event_threshold (specified at port setup time). A port will also fail to enqueue if it lacks enough hardware credits to enqueue; load-balanced credits are used to enqueue to a load-balanced queue, and directed credits are used to enqueue to a directed queue.

The out-of-credit situations are typically transient, and an eventdev application using the DLB ought to retry its enqueues if they fail. If enqueue fails, DLB PMD sets rte_errno as follows:

  • -ENOSPC: Credit exhaustion (either hardware or software)
  • -EINVAL: Invalid argument, such as port ID, queue ID, or sched_type.

Depending on the pipeline the application has constructed, it’s possible to enter a credit deadlock scenario wherein the worker thread lacks the credit to enqueue an event, and it must dequeue an event before it can recover the credit. If the worker thread retries its enqueue indefinitely, it will not make forward progress. Such deadlock is possible if the application has event “loops”, in which an event in dequeued from queue A and later enqueued back to queue A.

Due to this, workers should stop retrying after a time, release the events it is attempting to enqueue, and dequeue more events. It is important that the worker release the events and don’t simply set them aside to retry the enqueue again later, because the port has limited history list size (by default, twice the port’s dequeue_depth).

2.3.7. Priority

The DLB supports event priority and per-port queue service priority, as described in the eventdev header file. The DLB does not support ‘global’ event queue priority established at queue creation time.

DLB supports 4 event and queue service priority levels. For both priority types, the PMD uses the upper three bits of the priority field to determine the DLB priority, discarding the 5 least significant bits. But least significant bit out of 3 priority bits is effectively ignored for binning into 4 priorities. The discarded 5 least significant event priority bits are not preserved when an event is enqueued.

Note that event priority only works within the same event type. When atomic and ordered or unordered events are enqueued to same QID, priority across the types is always equal, and both types are served in a round robin manner.

2.3.8. Reconfiguration

The Eventdev API allows one to reconfigure a device, its ports, and its queues by first stopping the device, calling the configuration function(s), then restarting the device. The DLB does not support configuring an individual queue or port without first reconfiguring the entire device, however, so there are certain reconfiguration sequences that are valid in the eventdev API but not supported by the PMD.

Specifically, the PMD supports the following configuration sequence: 1. Configure and start the device 2. Stop the device 3. (Optional) Reconfigure the device 4. (Optional) If step 3 is run:

  1. Setup queue(s). The reconfigured queue(s) lose their previous port links.
  2. The reconfigured port(s) lose their previous queue links.
  1. (Optional, only if steps 4a and 4b are run) Link port(s) to queue(s)
  2. Restart the device. If the device is reconfigured in step 3 but one or more of its ports or queues are not, the PMD will apply their previous configuration (including port->queue links) at this time.

The PMD does not support the following configuration sequences: 1. Configure and start the device 2. Stop the device 3. Setup queue or setup port 4. Start the device

This sequence is not supported because the event device must be reconfigured before its ports or queues can be.

2.3.9. Atomic Inflights Allocation

In the last stage prior to scheduling an atomic event to a CQ, DLB holds the inflight event in a temporary buffer that is divided among load-balanced queues. If a queue’s atomic buffer storage fills up, this can result in head-of-line-blocking. For example:

  • An LDB queue allocated N atomic buffer entries
  • All N entries are filled with events from flow X, which is pinned to CQ 0.

Until CQ 0 releases 1+ events, no other atomic flows for that LDB queue can be scheduled. The likelihood of this case depends on the eventdev configuration, traffic behavior, event processing latency, potential for a worker to be interrupted or otherwise delayed, etc.

By default, the PMD allocates 16 buffer entries for each load-balanced queue, which provides an even division across all 128 queues but potentially wastes buffer space (e.g. if not all queues are used, or aren’t used for atomic scheduling).

The PMD provides a dev arg to override the default per-queue allocation. To increase per-queue atomic-inflight allocation to (for example) 64:

--allow ea:00.0,atm_inflights=64

2.3.10. QID Depth Threshold

DLB supports setting and tracking queue depth thresholds. Hardware uses the thresholds to track how full a queue is compared to its threshold. Four buckets are used

  • Less than or equal to 50% of queue depth threshold
  • Greater than 50%, but less than or equal to 75% of depth threshold
  • Greater than 75%, but less than or equal to 100% of depth threshold
  • Greater than 100% of depth thresholds

Per queue threshold metrics are tracked in the DLB xstats, and are also returned in the impl_opaque field of each received event.

The per qid threshold can be specified as part of the device args, and can be applied to all queue, a range of queues, or a single queue, as shown below.

--allow ea:00.0,qid_depth_thresh=all:<threshold_value>
--allow ea:00.0,qid_depth_thresh=qidA-qidB:<threshold_value>
--allow ea:00.0,qid_depth_thresh=qid:<threshold_value>

2.3.11. Class of service

DLB supports provisioning the DLB bandwidth into 4 classes of service.

  • Class 4 corresponds to 40% of the DLB hardware bandwidth
  • Class 3 corresponds to 30% of the DLB hardware bandwidth
  • Class 2 corresponds to 20% of the DLB hardware bandwidth
  • Class 1 corresponds to 10% of the DLB hardware bandwidth
  • Class 0 corresponds to don’t care

The classes are applied globally to the set of ports contained in this scheduling domain, which is more appropriate for the bifurcated PMD than for the PF PMD, since the PF PMD supports just 1 scheduling domain.

Class of service can be specified in the devargs, as follows

--allow ea:00.0,cos=<0..4>

2.3.12. Use X86 Vector Instructions

DLB supports using x86 vector instructions to optimize the data path.

The default mode of operation is to use scalar instructions, but the use of vector instructions can be enabled in the devargs, as follows

--allow ea:00.0,vector_opts_enabled=<y/Y>