Scaling in the Linux Networking Stack


Introduction
============

This document describes a set of complementary techniques in the Linux
networking stack to increase parallelism and improve performance for
multi-processor systems.

The following technologies are described:

  RSS: Receive Side Scaling
  RPS: Receive Packet Steering
  RFS: Receive Flow Steering
  Accelerated Receive Flow Steering
  XPS: Transmit Packet Steering


RSS: Receive Side Scaling
=========================

Contemporary NICs support multiple receive and transmit descriptor queues
(multi-queue). On reception, a NIC can send different packets to different
queues to distribute processing among CPUs. The NIC distributes packets by
applying a filter to each packet that assigns it to one of a small number
of logical flows. Packets for each flow are steered to a separate receive
queue, which in turn can be processed by separate CPUs. This mechanism is
generally known as “Receive-side Scaling” (RSS). The goal of RSS and
the other scaling techniques is to increase performance uniformly.
Multi-queue distribution can also be used for traffic prioritization, but
that is not the focus of these techniques.

The filter used in RSS is typically a hash function over the network
and/or transport layer headers-- for example, a 4-tuple hash over
IP addresses and TCP ports of a packet. The most common hardware
implementation of RSS uses a 128-entry indirection table where each entry
stores a queue number. The receive queue for a packet is determined
by masking out the low order seven bits of the computed hash for the
packet (usually a Toeplitz hash), taking this number as a key into the
indirection table and reading the corresponding value.

Some advanced NICs allow steering packets to queues based on
programmable filters. For example, webserver bound TCP port 80 packets
can be directed to their own receive queue. Such “n-tuple” filters can
be configured from ethtool (--config-ntuple).

==== RSS Configuration

The driver for a multi-queue capable NIC typically provides a kernel
module parameter for specifying the number of hardware queues to
configure. In the bnx2x driver, for instance, this parameter is called
num_queues. A typical RSS configuration would be to have one receive queue
for each CPU if the device supports enough queues, or otherwise at least
one for each memory domain, where a memory domain is a set of CPUs that
share a particular memory level (L1, L2, NUMA node, etc.).

The indirection table of an RSS device, which resolves a queue by masked
hash, is usually programmed by the driver at initialization. The
default mapping is to distribute the queues evenly in the table, but the
indirection table can be retrieved and modified at runtime using ethtool
commands (--show-rxfh-indir and --set-rxfh-indir). Modifying the
indirection table could be done to give different queues different
relative weights.

== RSS IRQ Configuration

Each receive queue has a separate IRQ associated with it. The NIC triggers
this to notify a CPU when new packets arrive on the given queue. The
signaling path for PCIe devices uses message signaled interrupts (MSI-X),
that can route each interrupt to a particular CPU. The active mapping
of queues to IRQs can be determined from /proc/interrupts. By default,
an IRQ may be handled on any CPU. Because a non-negligible part of packet
processing takes place in receive interrupt handling, it is advantageous
to spread receive interrupts between CPUs. To manually adjust the IRQ
affinity of each interrupt see Documentation/IRQ-affinity.txt. Some systems
will be running irqbalance, a daemon that dynamically optimizes IRQ
assignments and as a result may override any manual settings.

== Suggested Configuration

RSS should be enabled when latency is a concern or whenever receive
interrupt processing forms a bottleneck. Spreading load between CPUs
decreases queue length. For low latency networking, the optimal setting
is to allocate as many queues as there are CPUs in the system (or the
NIC maximum, if lower). The most efficient high-rate configuration
is likely the one with the smallest number of receive queues where no
receive queue overflows due to a saturated CPU, because in default
mode with interrupt coalescing enabled, the aggregate number of
interrupts (and thus work) grows with each additional queue.

Per-cpu load can be observed using the mpstat utility, but note that on
processors with hyperthreading (HT), each hyperthread is represented as
a separate CPU. For interrupt handling, HT has shown no benefit in
initial tests, so limit the number of queues to the number of CPU cores
in the system.


RPS: Receive Packet Steering
============================

Receive Packet Steering (RPS) is logically a software implementation of
RSS. Being in software, it is necessarily called later in the datapath.
Whereas RSS selects the queue and hence CPU that will run the hardware
interrupt handler, RPS selects the CPU to perform protocol processing
above the interrupt handler. This is accomplished by placing the packet
on the desired CPU’s backlog queue and waking up the CPU for processing.
RPS has some advantages over RSS: 1) it can be used with any NIC,
2) software filters can easily be added to hash over new protocols,
3) it does not increase hardware device interrupt rate (although it does
introduce inter-processor interrupts (IPIs)).

RPS is called during bottom half of the receive interrupt handler, when
a driver sends a packet up the network stack with netif_rx() or
netif_receive_skb(). These call the get_rps_cpu() function, which
selects the queue that should process a packet.

The first step in determining the target CPU for RPS is to calculate a
flow hash over the packet’s addresses or ports (2-tuple or 4-tuple hash
depending on the protocol). This serves as a consistent hash of the
associated flow of the packet. The hash is either provided by hardware
or will be computed in the stack. Capable hardware can pass the hash in
the receive descriptor for the packet; this would usually be the same
hash used for RSS (e.g. computed Toeplitz hash). The hash is saved in
skb->rx_hash and can be used elsewhere in the stack as a hash of the
packet’s flow.

Each receive hardware queue has an associated list of CPUs to which
RPS may enqueue packets for processing. For each received packet,
an index into the list is computed from the flow hash modulo the size
of the list. The indexed CPU is the target for processing the packet,
and the packet is queued to the tail of that CPU’s backlog queue. At
the end of the bottom half routine, IPIs are sent to any CPUs for which
packets have been queued to their backlog queue. The IPI wakes backlog
processing on the remote CPU, and any queued packets are then processed
up the networking stack.

==== RPS Configuration

RPS requires a kernel compiled with the CONFIG_RPS kconfig symbol (on
by default for SMP). Even when compiled in, RPS remains disabled until
explicitly configured. The list of CPUs to which RPS may forward traffic
can be configured for each receive queue using a sysfs file entry:

 /sys/class/net/<dev>/queues/rx-<n>/rps_cpus

This file implements a bitmap of CPUs. RPS is disabled when it is zero
(the default), in which case packets are processed on the interrupting
CPU. Documentation/IRQ-affinity.txt explains how CPUs are assigned to
the bitmap.

== Suggested Configuration

For a single queue device, a typical RPS configuration would be to set
the rps_cpus to the CPUs in the same memory domain of the interrupting
CPU. If NUMA locality is not an issue, this could also be all CPUs in
the system. At high interrupt rate, it might be wise to exclude the
interrupting CPU from the map since that already performs much work.

For a multi-queue system, if RSS is configured so that a hardware
receive queue is mapped to each CPU, then RPS is probably redundant
and unnecessary. If there are fewer hardware queues than CPUs, then
RPS might be beneficial if the rps_cpus for each queue are the ones that
share the same memory domain as the interrupting CPU for that queue.

==== RPS Flow Limit

RPS scales kernel receive processing across CPUs without introducing
reordering. The trade-off to sending all packets from the same flow
to the same CPU is CPU load imbalance if flows vary in packet rate.
In the extreme case a single flow dominates traffic. Especially on
common server workloads with many concurrent connections, such
behavior indicates a problem such as a misconfiguration or spoofed
source Denial of Service attack.

Flow Limit is an optional RPS feature that prioritizes small flows
during CPU contention by dropping packets from large flows slightly
ahead of those from small flows. It is active only when an RPS or RFS
destination CPU approaches saturation.  Once a CPU's input packet
queue exceeds half the maximum queue length (as set by sysctl
net.core.netdev_max_backlog), the kernel starts a per-flow packet
count over the last 256 packets. If a flow exceeds a set ratio (by
default, half) of these packets when a new packet arrives, then the
new packet is dropped. Packets from other flows are still only
dropped once the input packet queue reaches netdev_max_backlog.
No packets are dropped when the input packet queue length is below
the threshold, so flow limit does not sever connections outright:
even large flows maintain connectivity.

== Interface

Flow limit is compiled in by default (CONFIG_NET_FLOW_LIMIT), but not
turned on. It is implemented for each CPU independently (to avoid lock
and cache contention) and toggled per CPU by setting the relevant bit
in sysctl net.core.flow_limit_cpu_bitmap. It exposes the same CPU
bitmap interface as rps_cpus (see above) when called from procfs:

 /proc/sys/net/core/flow_limit_cpu_bitmap

Per-flow rate is calculated by hashing each packet into a hashtable
bucket and incrementing a per-bucket counter. The hash function is
the same that selects a CPU in RPS, but as the number of buckets can
be much larger than the number of CPUs, flow limit has finer-grained
identification of large flows and fewer false positives. The default
table has 4096 buckets. This value can be modified through sysctl

 net.core.flow_limit_table_len

The value is only consulted when a new table is allocated. Modifying
it does not update active tables.

== Suggested Configuration

Flow limit is useful on systems with many concurrent connections,
where a single connection taking up 50% of a CPU indicates a problem.
In such environments, enable the feature on all CPUs that handle
network rx interrupts (as set in /proc/irq/N/smp_affinity).

The feature depends on the input packet queue length to exceed
the flow limit threshold (50%) + the flow history length (256).
Setting net.core.netdev_max_backlog to either 1000 or 10000
performed well in experiments.


RFS: Receive Flow Steering
==========================

While RPS steers packets solely based on hash, and thus generally
provides good load distribution, it does not take into account
application locality. This is accomplished by Receive Flow Steering
(RFS). The goal of RFS is to increase datacache hitrate by steering
kernel processing of packets to the CPU where the application thread
consuming the packet is running. RFS relies on the same RPS mechanisms
to enqueue packets onto the backlog of another CPU and to wake up that
CPU.

In RFS, packets are not forwarded directly by the value of their hash,
but the hash is used as index into a flow lookup table. This table maps
flows to the CPUs where those flows are being processed. The flow hash
(see RPS section above) is used to calculate the index into this table.
The CPU recorded in each entry is the one which last processed the flow.
If an entry does not hold a valid CPU, then packets mapped to that entry
are steered using plain RPS. Multiple table entries may point to the
same CPU. Indeed, with many flows and few CPUs, it is very likely that
a single application thread handles flows with many different flow hashes.

rps_sock_flow_table is a global flow table that contains the *desired* CPU
for flows: the CPU that is currently processing the flow in userspace.
Each table value is a CPU index that is updated during calls to recvmsg
and sendmsg (specifically, inet_recvmsg(), inet_sendmsg(), inet_sendpage()
and tcp_splice_read()).

When the scheduler moves a thread to a new CPU while it has outstanding
receive packets on the old CPU, packets may arrive out of order. To
avoid this, RFS uses a second flow table to track outstanding packets
for each flow: rps_dev_flow_table is a table specific to each hardware
receive queue of each device. Each table value stores a CPU index and a
counter. The CPU index represents the *current* CPU onto which packets
for this flow are enqueued for further kernel processing. Ideally, kernel
and userspace processing occur on the same CPU, and hence the CPU index
in both tables is identical. This is likely false if the scheduler has
recently migrated a userspace thread while the kernel still has packets
enqueued for kernel processing on the old CPU.

The counter in rps_dev_flow_table values records the length of the current
CPU's backlog when a packet in this flow was last enqueued. Each backlog
queue has a head counter that is incremented on dequeue. A tail counter
is computed as head counter + queue length. In other words, the counter
in rps_dev_flow[i] records the last element in flow i that has
been enqueued onto the currently designated CPU for flow i (of course,
entry i is actually selected by hash and multiple flows may hash to the
same entry i).

And now the trick for avoiding out of order packets: when selecting the
CPU for packet processing (from get_rps_cpu()) the rps_sock_flow table
and the rps_dev_flow table of the queue that the packet was received on
are compared. If the desired CPU for the flow (found in the
rps_sock_flow table) matches the current CPU (found in the rps_dev_flow
table), the packet is enqueued onto that CPU’s backlog. If they differ,
the current CPU is updated to match the desired CPU if one of the
following is true:

- The current CPU's queue head counter >= the recorded tail counter
  value in rps_dev_flow[i]
- The current CPU is unset (equal to RPS_NO_CPU)
- The current CPU is offline

After this check, the packet is sent to the (possibly updated) current
CPU. These rules aim to ensure that a flow only moves to a new CPU when
there are no packets outstanding on the old CPU, as the outstanding
packets could arrive later than those about to be processed on the new
CPU.

==== RFS Configuration

RFS is only available if the kconfig symbol CONFIG_RPS is enabled (on
by default for SMP). The functionality remains disabled until explicitly
configured. The number of entries in the global flow table is set through:

 /proc/sys/net/core/rps_sock_flow_entries

The number of entries in the per-queue flow table are set through:

 /sys/class/net/<dev>/queues/rx-<n>/rps_flow_cnt

== Suggested Configuration

Both of these need to be set before RFS is enabled for a receive queue.
Values for both are rounded up to the nearest power of two. The
suggested flow count depends on the expected number of active connections
at any given time, which may be significantly less than the number of open
connections. We have found that a value of 32768 for rps_sock_flow_entries
works fairly well on a moderately loaded server.

For a single queue device, the rps_flow_cnt value for the single queue
would normally be configured to the same value as rps_sock_flow_entries.
For a multi-queue device, the rps_flow_cnt for each queue might be
configured as rps_sock_flow_entries / N, where N is the number of
queues. So for instance, if rps_sock_flow_entries is set to 32768 and there
are 16 configured receive queues, rps_flow_cnt for each queue might be
configured as 2048.


Accelerated RFS
===============

Accelerated RFS is to RFS what RSS is to RPS: a hardware-accelerated load
balancing mechanism that uses soft state to steer flows based on where
the application thread consuming the packets of each flow is running.
Accelerated RFS should perform better than RFS since packets are sent
directly to a CPU local to the thread consuming the data. The target CPU
will either be the same CPU where the application runs, or at least a CPU
which is local to the application thread’s CPU in the cache hierarchy.

To enable accelerated RFS, the networking stack calls the
ndo_rx_flow_steer driver function to communicate the desired hardware
queue for packets matching a particular flow. The network stack
automatically calls this function every time a flow entry in
rps_dev_flow_table is updated. The driver in turn uses a device specific
method to program the NIC to steer the packets.

The hardware queue for a flow is derived from the CPU recorded in
rps_dev_flow_table. The stack consults a CPU to hardware queue map which
is maintained by the NIC driver. This is an auto-generated reverse map of
the IRQ affinity table shown by /proc/interrupts. Drivers can use
functions in the cpu_rmap (“CPU affinity reverse map”) kernel library
to populate the map. For each CPU, the corresponding queue in the map is
set to be one whose processing CPU is closest in cache locality.

==== Accelerated RFS Configuration

Accelerated RFS is only available if the kernel is compiled with
CONFIG_RFS_ACCEL and support is provided by the NIC device and driver.
It also requires that ntuple filtering is enabled via ethtool. The map
of CPU to queues is automatically deduced from the IRQ affinities
configured for each receive queue by the driver, so no additional
configuration should be necessary.

== Suggested Configuration

This technique should be enabled whenever one wants to use RFS and the
NIC supports hardware acceleration.

XPS: Transmit Packet Steering
=============================

Transmit Packet Steering is a mechanism for intelligently selecting
which transmit queue to use when transmitting a packet on a multi-queue
device. To accomplish this, a mapping from CPU to hardware queue(s) is
recorded. The goal of this mapping is usually to assign queues
exclusively to a subset of CPUs, where the transmit completions for
these queues are processed on a CPU within this set. This choice
provides two benefits. First, contention on the device queue lock is
significantly reduced since fewer CPUs contend for the same queue
(contention can be eliminated completely if each CPU has its own
transmit queue). Secondly, cache miss rate on transmit completion is
reduced, in particular for data cache lines that hold the sk_buff
structures.

XPS is configured per transmit queue by setting a bitmap of CPUs that
may use that queue to transmit. The reverse mapping, from CPUs to
transmit queues, is computed and maintained for each network device.
When transmitting the first packet in a flow, the function
get_xps_queue() is called to select a queue. This function uses the ID
of the running CPU as a key into the CPU-to-queue lookup table. If the
ID matches a single queue, that is used for transmission. If multiple
queues match, one is selected by using the flow hash to compute an index
into the set.

The queue chosen for transmitting a particular flow is saved in the
corresponding socket structure for the flow (e.g. a TCP connection).
This transmit queue is used for subsequent packets sent on the flow to
prevent out of order (ooo) packets. The choice also amortizes the cost
of calling get_xps_queues() over all packets in the flow. To avoid
ooo packets, the queue for a flow can subsequently only be changed if
skb->ooo_okay is set for a packet in the flow. This flag indicates that
there are no outstanding packets in the flow, so the transmit queue can
change without the risk of generating out of order packets. The
transport layer is responsible for setting ooo_okay appropriately. TCP,
for instance, sets the flag when all data for a connection has been
acknowledged.

==== XPS Configuration

XPS is only available if the kconfig symbol CONFIG_XPS is enabled (on by
default for SMP). The functionality remains disabled until explicitly
configured. To enable XPS, the bitmap of CPUs that may use a transmit
queue is configured using the sysfs file entry:

/sys/class/net/<dev>/queues/tx-<n>/xps_cpus

== Suggested Configuration

For a network device with a single transmission queue, XPS configuration
has no effect, since there is no choice in this case. In a multi-queue
system, XPS is preferably configured so that each CPU maps onto one queue.
If there are as many queues as there are CPUs in the system, then each
queue can also map onto one CPU, resulting in exclusive pairings that
experience no contention. If there are fewer queues than CPUs, then the
best CPUs to share a given queue are probably those that share the cache
with the CPU that processes transmit completions for that queue
(transmit interrupts).


Further Information
===================
RPS and RFS were introduced in kernel 2.6.35. XPS was incorporated into
2.6.38. Original patches were submitted by Tom Herbert
(therbert@google.com)

Accelerated RFS was introduced in 2.6.35. Original patches were
submitted by Ben Hutchings (bhutchings@solarflare.com)

Authors:
Tom Herbert (therbert@google.com)
Willem de Bruijn (willemb@google.com)