[lkml]   [2016]   [Aug]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[RFC V2 PATCH 00/25] Kernel NET policy
From: Kan Liang <>

(re-send to correct system time issue. Sorry for any inconvenience.)
It is a big challenge to get good network performance. First, the network
performance is not good with default system settings. Second, it is too
difficult to do automatic tuning for all possible workloads, since workloads
have different requirements. Some workloads may want high throughput. Some may
need low latency. Last but not least, there are lots of manual configurations.
Fine grained configuration is too difficult for users.

NET policy intends to simplify the network configuration and get a good network
performance according to the hints(policy) which is applied by user. It
provides some typical "policies" for user which can be set per-socket, per-task
or per-device. The kernel will automatically figures out how to merge different
requests to get good network performance.

NET policy is designed for multiqueue network devices. This implementation is
only for Intel NICs using i40e driver. But the concepts and generic code should
apply to other multiqueue NICs too.

NET policy is also a combination of generic policy manager code and some
ethtool callbacks (per queue coalesce setting, flow classification rules) to
configure the driver.

This series also supports CPU hotplug and device hotplug.

Here are some common questions about NET policy.
1. Why userspace tool cannot do the same thing?
A: Kernel is more suitable for NET policy.
- User space code would be far more complicated to get right and perform
well . It always need to work with out of date state compared to the
latest, because it cannot do any locking with the kernel state.
- User space code is less efficient than kernel code, because of the
additional context switches needed.
- Kernel is in the right position to coordinate requests from multiple

2. Is NET policy looking for optimal settings?
A: No. The NET policy intends to get a good network performance according
to user's specific request. Our target for good performance is ~90% of
the optimal settings.

3. How's the configuration impact the connection rates?
A: There are two places to acquire rtnl mutex to configure the device.
- One is to do device policy setting. It happens on initalization stage,
hotplug or queue number changes. The device policy will be set to
NET_POLICY_NONE. If so, it "falls back" to the system default way to
direct the packets. It doesn't block the connection.
- The other is to set Rx network flow classification options or rules.
It uses work queue to do asynchronized setting. It avoid destroying
the connection rates.

4. Why not using existing mechanism for NET policy?
For example, cgroup tc or existing SOCKET options.
A: The NET policy has already used existing mechanism as many as it can.
For example, it uses existing ethtool interface to configure the device.
However, the NET policy stiil need to introduce new interfaces to meet
its special request.
For resource usage, current cgroup tc is not suitable for per-socket
setting. Also, current tc can only set rate limit. The NET policy wants
to change interrupt moderation per device queue. So in this series, it
will not use cgroup tc. But in some places, cgroup and NET policy are
similar. For example, both of them isolates the resource usage. Both of
them do traffic controller. So it is on the NET policy TODO list to
work well with cgroup.
For socket options, SO_MARK or may be SO_PRIORITY is close to NET policy's
requirement. But they can not be reused for NET policy. SO_MARK can be
used for routing and packet filtering. But the NET policy doesn't intend to
change the routing. It only redirects the packet to the specific device
queue. Also, the target queue is assigned by NET policy subsystem at run
time. It should not be set in advance. SO_PRIORITY can set protocol-defined
priority for all packets on the socket. But the policies don't have priority.

5. Why disable IRQ balance?
A: Disabling IRQ balance is a common way (recommend way for some devices) to
tune network performance.

Here are some key Interfaces/APIs for NET policy.

Interfaces which export to user space

User can set/get per device policy from /proc

User can set/get per task policy from /proc
An alternative way to set/get per task policy is from prctl.

User can set/get per socket policy by setsockopt

New ndo opt

int (*ndo_netpolicy_init)(struct net_device *dev,
struct netpolicy_info *info);
Initialize device driver for NET policy

int (*ndo_get_irq_info)(struct net_device *dev,
struct netpolicy_dev_info *info);
Collect device information. Currently, only collecting IRQ
informance should be enough.

int (*ndo_set_net_policy)(struct net_device *dev,
enum netpolicy_name name);
This interface is used to set device NET policy by name. It is device driver's
responsibility to set driver specific configuration for the given policy.

NET policy subsystem APIs

netpolicy_register(struct netpolicy_instance *instance,
enum netpolicy_name policy)
netpolicy_unregister(struct netpolicy_instance *instance)
Register/unregister per task/socket NET policy.
The socket/task can only be benefited when it register itself with
specific policy. After registeration, an record will be created and inserted
into a RCU hash table, which include all the NET policy related information
for the socket/task.

netpolicy_pick_queue(struct netpolicy_instance *instance, bool is_rx);
Find the proper queue according to policy for packet receiving and

netpolicy_set_rules(struct netpolicy_instance *instance);
Configure Rx network flow classification rules

For using NET policy, the per-device policy must be set in advance. It will
automatically configure the system and re-organize the resource of the system
accordingly. For system configuration, in this series, it will disable irq
balance, set device queue irq affinity, and modify interrupt moderation. For
re-organizing the resource, current implementation forces that CPU and queue
irq are 1:1 mapping. An 1:1 mapping group is also called NET policy object.
For each device policy, it maintains a policy list. Once the device policy is
applied, the objects will be insert and tracked in that device policy list. The
policy list only be updated when CPU/device hotplug, queue number changes or
device policy changes.
The user can use /proc, prctl and setsockopt to set per-task and per-socket
NET policy. Once the policy is set, an related record will be inserted into RCU
hash table. The record includes ptr, policy and NET policy object. The ptr is
the pointer address of task/socket. The object will not be assigned until the
first package receive/transmit. The object is picked by round-robin from object
list. Once the object is determined, the following packets will be set to
redirect to the queue(object).
The object can be shared. The per-task or per-socket policy can be inherited.

Now NET policy supports four per device policies and three per task/socket
- BULK policy: This policy is designed for high throughput. It can be
applied to either per device policy or per task/socket policy.
- CPU policy: This policy is designed for high throughput but lower CPU
utilization (power saving). It can be applied to either per device policy
or per task/socket policy.
- LATENCY policy: This policy is designed for low latency. It can be
applied to either per device policy or per task/socket policy.
- MIX policy: This policy can only be applied to per device policy. This
is designed for the case which miscellaneous types of workload running
on the device.

Lots of tests are done for NET policy on platforms with Intel Xeon E5 V2
and XL710 40G NIC. The baseline test is with Linux 4.6.0 kernel.
Netperf is used to evaluate the throughput and latency performance.
- "netperf -f m -t TCP_RR -H server_IP -c -C -l 60 -- -r buffersize
-b burst -D" is used to evaluate throughput performance, which is
called throughput-first workload.
- "netperf -t TCP_RR -H server_IP -c -C -l 60 -- -r buffersize" is
used to evaluate latency performance, which is called latency-first
- Different loads are also evaluated by running 1, 12, 24, 48 or 96
throughput-first workloads/latency-first workload simultaneously.

For "BULK" policy, the throughput performance is on average ~1.27X than
For "CPU" policy, the throughput performance is on average ~1.25X than
baseline, and has lower CPU% (on average ~5% lower than "BULK" policy).
For "LATENCY" policy, the latency is on average 51.5% less than the baseline.
For "MIX" policy, mixed workloads performance is evaluated.
The mixed workloads are combination of throughput-first workload and
latency-first workload. Five different types of combinations are evaluated
(pure throughput-first workload, pure latency-first workloads,
2/3 throughput-first workload + 1/3 latency-first workloads,
1/3 throughput-first workload + 2/3 latency-first workloads and
1/2 throughput-first workload + 1/2 latency-first workloads).
For caculating the performance of mixed workloads, a weighted sum system
is introduced.
Score = normalized_latency * Weight + normalized_throughput * (1 - Weight).
If we assume that the user has an equal interest in latency and throughput
performance, the Score for "MIX" policy is on average ~1.83X than baseline.

Changes since V1:
- Using work queue to set Rx network flow classification rules and search
available NET policy object asynchronously.
- Using RCU lock to replace read-write lock
- Redo performance test and update performance results.
- Some minor modification for codes and documents.
- Remove i40e related patches which will be submitted in separate thread.

Kan Liang (25):
net: introduce NET policy
net/netpolicy: init NET policy
net/netpolicy: get device queue irq information
net/netpolicy: get CPU information
net/netpolicy: create CPU and queue mapping
net/netpolicy: set and remove IRQ affinity
net/netpolicy: enable and disable NET policy
net/netpolicy: introduce NET policy object
net/netpolicy: set NET policy by policy name
net/netpolicy: add three new NET policies
net/netpolicy: add MIX policy
net/netpolicy: NET device hotplug
net/netpolicy: support CPU hotplug
net/netpolicy: handle channel changes
net/netpolicy: implement netpolicy register
net/netpolicy: introduce per socket netpolicy
net/netpolicy: introduce netpolicy_pick_queue
net/netpolicy: set tx queues according to policy
net/netpolicy: set Rx queues according to policy
net/netpolicy: introduce per task net policy
net/netpolicy: set per task policy by proc
net/netpolicy: fast path for finding the queues
net/netpolicy: optimize for queue pair
net/netpolicy: limit the total record number
Documentation/networking: Document NET policy

Documentation/networking/netpolicy.txt | 157 ++++
arch/alpha/include/uapi/asm/socket.h | 2 +
arch/avr32/include/uapi/asm/socket.h | 2 +
arch/frv/include/uapi/asm/socket.h | 2 +
arch/ia64/include/uapi/asm/socket.h | 2 +
arch/m32r/include/uapi/asm/socket.h | 2 +
arch/mips/include/uapi/asm/socket.h | 2 +
arch/mn10300/include/uapi/asm/socket.h | 2 +
arch/parisc/include/uapi/asm/socket.h | 2 +
arch/powerpc/include/uapi/asm/socket.h | 2 +
arch/s390/include/uapi/asm/socket.h | 2 +
arch/sparc/include/uapi/asm/socket.h | 2 +
arch/xtensa/include/uapi/asm/socket.h | 2 +
fs/proc/base.c | 64 ++
include/linux/init_task.h | 9 +
include/linux/netdevice.h | 31 +
include/linux/netpolicy.h | 163 ++++
include/linux/sched.h | 5 +
include/net/net_namespace.h | 3 +
include/net/request_sock.h | 4 +-
include/net/sock.h | 28 +
include/uapi/asm-generic/socket.h | 2 +
include/uapi/linux/prctl.h | 4 +
kernel/exit.c | 4 +
kernel/fork.c | 6 +
kernel/sys.c | 31 +
net/Kconfig | 7 +
net/core/Makefile | 1 +
net/core/dev.c | 20 +-
net/core/ethtool.c | 8 +-
net/core/netpolicy.c | 1511 ++++++++++++++++++++++++++++++++
net/core/sock.c | 36 +
net/ipv4/af_inet.c | 71 ++
net/ipv4/udp.c | 4 +
34 files changed, 2189 insertions(+), 4 deletions(-)
create mode 100644 Documentation/networking/netpolicy.txt
create mode 100644 include/linux/netpolicy.h
create mode 100644 net/core/netpolicy.c


 \ /
  Last update: 2016-08-04 22:21    [W:0.086 / U:0.084 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site