Sleep for an incremental amount of time if none of the Rx queues
assigned to a PMD have at least half a batch of packets (i.e. 16 pkts)
on an polling iteration of the PMD.
Upon detecting the threshold of >= 16 pkts on an Rxq, reset the
sleep time to zero (i.e. no sleep).
Sleep time will be increased on each iteration where the low load
conditions remain up to a total of the max sleep time which is set
by the user e.g:
ovs-vsctl set Open_vSwitch . other_config:pmd-maxsleep=500
The default pmd-maxsleep value is 0, which means that no sleeps
will occur and the default behaviour is unchanged from previously.
Also add new stats to pmd-perf-show to get visibility of operation
e.g.
...
- sleep iterations: 153994 ( 76.8 % of iterations)
Sleep time (us): 9159399 ( 59 us/iteration avg.)
...
Reviewed-by: Robin Jarry <rjarry@redhat.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Signed-off-by: Kevin Traynor <ktraynor@redhat.com>
Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
There are cases where users might want simple forwarding or drop rules
for all packets received from a specific port, e.g ::
"in_port=1,actions=2"
"in_port=2,actions=IN_PORT"
"in_port=3,vlan_tci=0x1234/0x1fff,actions=drop"
"in_port=4,actions=push_vlan:0x8100,set_field:4196->vlan_vid,output:3"
There are also cases where complex OpenFlow rules can be simplified
down to datapath flows with very simple match criteria.
In theory, for very simple forwarding, OVS doesn't need to parse
packets at all in order to follow these rules. "Simple match" lookup
optimization is intended to speed up packet forwarding in these cases.
Design:
Due to various implementation constraints userspace datapath has
following flow fields always in exact match (i.e. it's required to
match at least these fields of a packet even if the OF rule doesn't
need that):
- recirc_id
- in_port
- packet_type
- dl_type
- vlan_tci (CFI + VID) - in most cases
- nw_frag - for ip packets
Not all of these fields are related to packet itself. We already
know the current 'recirc_id' and the 'in_port' before starting the
packet processing. It also seems safe to assume that we're working
with Ethernet packets. So, for the simple OF rule we need to match
only on 'dl_type', 'vlan_tci' and 'nw_frag'.
'in_port', 'dl_type', 'nw_frag' and 13 bits of 'vlan_tci' can be
combined in a single 64bit integer (mark) that can be used as a
hash in hash map. We are using only VID and CFI form the 'vlan_tci',
flows that need to match on PCP will not qualify for the optimization.
Workaround for matching on non-existence of vlan updated to match on
CFI and VID only in order to qualify for the optimization. CFI is
always set by OVS if vlan is present in a packet, so there is no need
to match on PCP in this case. 'nw_frag' takes 2 bits of PCP inside
the simple match mark.
New per-PMD flow table 'simple_match_table' introduced to store
simple match flows only. 'dp_netdev_flow_add' adds flow to the
usual 'flow_table' and to the 'simple_match_table' if the flow
meets following constraints:
- 'recirc_id' in flow match is 0.
- 'packet_type' in flow match is Ethernet.
- Flow wildcards contains only minimal set of non-wildcarded fields
(listed above).
If the number of flows for current 'in_port' in a regular 'flow_table'
equals number of flows for current 'in_port' in a 'simple_match_table',
we may use simple match optimization, because all the flows we have
are simple match flows. This means that we only need to parse
'dl_type', 'vlan_tci' and 'nw_frag' to perform packet matching.
Now we make the unique flow mark from the 'in_port', 'dl_type',
'nw_frag' and 'vlan_tci' and looking for it in the 'simple_match_table'.
On successful lookup we don't need to run full 'miniflow_extract()'.
Unsuccessful lookup technically means that we have no suitable flow
in the datapath and upcall will be required. So, in this case EMC and
SMC lookups are disabled. We may optimize this path in the future by
bypassing the dpcls lookup too.
Performance improvement of this solution on a 'simple match' flows
should be comparable with partial HW offloading, because it parses same
packet fields and uses similar flow lookup scheme.
However, unlike partial HW offloading, it works for all port types
including virtual ones.
Performance results when compared to EMC:
Test setup:
virtio-user OVS virtio-user
Testpmd1 ------------> pmd1 ------------> Testpmd2
(txonly) x<------ pmd2 <------------ (mac swap)
Single stream of 64byte packets. Actions:
in_port=vhost0,actions=vhost1
in_port=vhost1,actions=vhost0
Stats collected from pmd1 and pmd2, so there are 2 scenarios:
Virt-to-Virt : Testpmd1 ------> pmd1 ------> Testpmd2.
Virt-to-NoCopy : Testpmd2 ------> pmd2 --->x Testpmd1.
Here the packet sent from pmd2 to Testpmd1 is always dropped, because
the virtqueue is full since Testpmd1 is in txonly mode and doesn't
receive any packets. This should be closer to the performance of a
VM-to-Phy scenario.
Test performed on machine with Intel Xeon CPU E5-2690 v4 @ 2.60GHz.
Table below represents improvement in throughput when compared to EMC.
+----------------+------------------------+------------------------+
| | Default (-g -O2) | "-Ofast -march=native" |
| Scenario +------------+-----------+------------+-----------+
| | GCC | Clang | GCC | Clang |
+----------------+------------+-----------+------------+-----------+
| Virt-to-Virt | +18.9% | +25.5% | +10.8% | +16.7% |
| Virt-to-NoCopy | +24.3% | +33.7% | +14.9% | +22.0% |
+----------------+------------+-----------+------------+-----------+
For Phy-to-Phy case performance improvement should be even higher, but
it's not the main use-case for this functionality. Performance
difference for the non-simple flows is within a margin of error.
Acked-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
This commit adds a new counter to be displayed to the user when
requesting datapath packet statistics. It counts the number of
packets that are parsed and a miniflow built up from it by the
optimized miniflow extract parsers.
The ovs-appctl command "dpif-netdev/pmd-perf-show" now has an
extra entry indicating if the optimized MFEX was hit:
- MFEX Opt hits: 6786432 (100.0 %)
Signed-off-by: Harry van Haaren <harry.van.haaren@intel.com>
Acked-by: Flavio Leitner <fbl@sysclose.org>
Acked-by: Eelco Chaudron <echaudro@redhat.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
It is possible for packets traversing the userspace datapath to match a
flow before hitting on EMC by using a mark ID provided by a NIC. Add a
PMD statistic for this hit.
Signed-off-by: Cian Ferriter <cian.ferriter@intel.com>
Acked-by: Flavio Leitner <fbl@sysclose.org>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
The accurate timing implementation in this patch gets the wall clock counter via
cntvct_el0 register access. This call is portable to all aarch64 architectures
and has been verified on an 64-bit arm server.
Suggested-by: Yanqin Wei <yanqin.wei@arm.com>
Reviewed-by: Ilya Maximets <i.maximets@ovn.org>
Signed-off-by: Malvika Gupta <malvika.gupta@arm.com>
Signed-off-by: Ben Pfaff <blp@ovn.org>
Unlike 'rte_get_tsc_cycles()' which doesn't need any specific
initialization, 'rte_get_tsc_hz()' could be used only after successfull
call to 'rte_eal_init()'. 'rte_eal_init()' estimates the TSC frequency
for later use by 'rte_get_tsc_hz()'. Fairly said, we're not allowed
to use 'rte_get_tsc_cycles()' before initializing DPDK too, but it
works this way for now and provides correct results.
This patch provides TSC frequency estimation code that will be used
in two cases:
* DPDK is not compiled in, i.e. DPDK_NETDEV not defined.
* DPDK compiled in but not initialized,
i.e. other_config:dpdk-init=false
This change is mostly useful for AF_XDP netdev support, i.e. allows
to use dpif-netdev/pmd-perf-show command and various PMD perf metrics.
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
Acked-by: William Tu <u9012063@gmail.com>
The patch introduces experimental AF_XDP support for OVS netdev.
AF_XDP, the Address Family of the eXpress Data Path, is a new Linux socket
type built upon the eBPF and XDP technology. It is aims to have comparable
performance to DPDK but cooperate better with existing kernel's networking
stack. An AF_XDP socket receives and sends packets from an eBPF/XDP program
attached to the netdev, by-passing a couple of Linux kernel's subsystems
As a result, AF_XDP socket shows much better performance than AF_PACKET
For more details about AF_XDP, please see linux kernel's
Documentation/networking/af_xdp.rst. Note that by default, this feature is
not compiled in.
Signed-off-by: William Tu <u9012063@gmail.com>
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Printing of the SMC hits missed in the 'dpif-netdev/pmd-perf-show'
appctl command.
CC: Yipeng Wang <yipeng1.wang@intel.com>
Fixes: 60d8ccae135f ("dpif-netdev: Add SMC cache after EMC cache")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Yipeng Wang <yipeng1.wang@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This patch adds a signature match cache (SMC) after exact match
cache (EMC). The difference between SMC and EMC is SMC only stores
a signature of a flow thus it is much more memory efficient. With
same memory space, EMC can store 8k flows while SMC can store 1M
flows. It is generally beneficial to turn on SMC but turn off EMC
when traffic flow count is much larger than EMC size.
SMC cache will map a signature to an dp_netdev_flow index in
flow_table. Thus, we add two new APIs in cmap for lookup key by
index and lookup index by key.
For now, SMC is an experimental feature that it is turned off by
default. One can turn it on using ovsdb options.
Signed-off-by: Yipeng Wang <yipeng1.wang@intel.com>
Co-authored-by: Jan Scheurich <jan.scheurich@ericsson.com>
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This patch enhances dpif-netdev-perf to detect iterations with
suspicious statistics according to the following criteria:
- iteration lasts longer than US_THR microseconds (default 250).
This can be used to capture events where a PMD is blocked or
interrupted for such a period of time that there is a risk for
dropped packets on any of its Rx queues.
- max vhost qlen exceeds a threshold Q_THR (default 128). This can
be used to infer virtio queue overruns and dropped packets inside
a VM, which are not visible in OVS otherwise.
Such suspicious iterations can be logged together with their iteration
statistics to be able to correlate them to packet drop or other events
outside OVS.
A new command is introduced to enable/disable logging at run-time and
to adjust the above thresholds for suspicious iterations:
ovs-appctl dpif-netdev/pmd-perf-log-set on | off
[-b before] [-a after] [-e|-ne] [-us usec] [-q qlen]
Turn logging on or off at run-time (on|off).
-b before: The number of iterations before the suspicious iteration to
be logged (default 5).
-a after: The number of iterations after the suspicious iteration to
be logged (default 5).
-e: Extend logging interval if another suspicious iteration is
detected before logging occurs.
-ne: Do not extend logging interval (default).
-q qlen: Suspicious vhost queue fill level threshold. Increase this
to 512 if the Qemu supports 1024 virtio queue length.
(default 128).
-us usec: change the duration threshold for a suspicious iteration
(default 250 us).
Note: Logging of suspicious iterations itself consumes a considerable
amount of processing cycles of a PMD which may be visible in the iteration
history. In the worst case this can lead OVS to detect another
suspicious iteration caused by logging.
If more than 100 iterations around a suspicious iteration have been
logged once, OVS falls back to the safe default values (-b 5/-a 5/-ne)
to avoid that logging itself causes continuos further logging.
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
This patch instruments the dpif-netdev datapath to record detailed
statistics of what is happening in every iteration of a PMD thread.
The collection of detailed statistics can be controlled by a new
Open_vSwitch configuration parameter "other_config:pmd-perf-metrics".
By default it is disabled. The run-time overhead, when enabled, is
in the order of 1%.
The covered metrics per iteration are:
- cycles
- packets
- (rx) batches
- packets/batch
- max. vhostuser qlen
- upcalls
- cycles spent in upcalls
This raw recorded data is used threefold:
1. In histograms for each of the following metrics:
- cycles/iteration (log.)
- packets/iteration (log.)
- cycles/packet
- packets/batch
- max. vhostuser qlen (log.)
- upcalls
- cycles/upcall (log)
The histograms bins are divided linear or logarithmic.
2. A cyclic history of the above statistics for 999 iterations
3. A cyclic history of the cummulative/average values per millisecond
wall clock for the last 1000 milliseconds:
- number of iterations
- avg. cycles/iteration
- packets (Kpps)
- avg. packets/batch
- avg. max vhost qlen
- upcalls
- avg. cycles/upcall
The gathered performance metrics can be printed at any time with the
new CLI command
ovs-appctl dpif-netdev/pmd-perf-show [-nh] [-it iter_len] [-ms ms_len]
[-pmd core] [dp]
The options are
-nh: Suppress the histograms
-it iter_len: Display the last iter_len iteration stats
-ms ms_len: Display the last ms_len millisecond stats
-pmd core: Display only the specified PMD
The performance statistics are reset with the existing
dpif-netdev/pmd-stats-clear command.
The output always contains the following global PMD statistics,
similar to the pmd-stats-show command:
Time: 15:24:55.270
Measurement duration: 1.008 s
pmd thread numa_id 0 core_id 1:
Cycles: 2419034712 (2.40 GHz)
Iterations: 572817 (1.76 us/it)
- idle: 486808 (15.9 % cycles)
- busy: 86009 (84.1 % cycles)
Rx packets: 2399607 (2381 Kpps, 848 cycles/pkt)
Datapath passes: 3599415 (1.50 passes/pkt)
- EMC hits: 336472 ( 9.3 %)
- Megaflow hits: 3262943 (90.7 %, 1.00 subtbl lookups/hit)
- Upcalls: 0 ( 0.0 %, 0.0 us/upcall)
- Lost upcalls: 0 ( 0.0 %)
Tx packets: 2399607 (2381 Kpps)
Tx batches: 171400 (14.00 pkts/batch)
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off-by: Ian Stokes <ian.stokes@intel.com>
Simplify the historically grown TSC cycle counting in PMD threads.
Cycles are currently counted for the following purposes:
1. Measure PMD ustilization
PMD utilization is defined as ratio of cycles spent in busy iterations
(at least one packet received or sent) over the total number of cycles.
This is already done in pmd_perf_start_iteration() and
pmd_perf_end_iteration() based on a TSC timestamp saved in current
iteration at start_iteration() and the actual TSC at end_iteration().
No dependency on intermediate cycle accounting.
2. Measure the processing load per RX queue
This comprises cycles spend on polling and processing packets received
from the rx queue and the cycles spent on delayed sending of these packets
to tx queues (with time-based batching).
The previous scheme using cycles_count_start(), cycles_count_intermediate()
and cycles-count_end() originally introduced to simplify cycle counting
and saving calls to rte_get_tsc_cycles() was rather obscuring things.
Replace by a nestable cycle_timer with with start and stop functions to
embrace a code segment to be timed. The timed code may contain arbitrary
nested cycle_timers. The duration of nested timers is excluded from the
outer timer.
The caller must ensure that each call to cycle_timer_start() is
followed by a call to cycle_timer_end(). Failure to do so will lead to
assertion failure or a memory leak.
The new cycle_timer is used to measure the processing cycles per rx queue.
This is not yet strictly necessary but will be made use of in a subsequent
commit.
All cycle count functions and data are relocated to module
dpif-netdev-perf.
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Acked-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off: Ian Stokes <ian.stokes@intel.com>
Add module dpif-netdev-perf to host all PMD performance-related
data structures and functions in dpif-netdev. Refactor the PMD
stats handling in dpif-netdev and delegate whatever possible into
the new module, using clean interfaces to shield dpif-netdev from
the implementation details. Accordingly, the all PMD statistics
members are moved from the main struct dp_netdev_pmd_thread into
a dedicated member of type struct pmd_perf_stats.
Include Darrel's prior refactoring of PMD stats contained in
[PATCH v5,2/3] dpif-netdev: Refactor some pmd stats:
1. The cycles per packet counts are now based on packets
received rather than packet passes through the datapath.
2. Packet counters are now kept for packets received and
packets recirculated. These are kept as separate counters for
maintainability reasons. The cost of incrementing these counters
is negligible. These new counters are also displayed to the user.
3. A display statistic is added for the average number of
datapath passes per packet. This should be useful for user
debugging and understanding of packet processing.
4. The user visible 'miss' counter is used for successful upcalls,
rather than the sum of sucessful and unsuccessful upcalls. Hence,
this becomes what user historically understands by OVS 'miss upcall'.
The user display is annotated to make this clear as well.
5. The user visible 'lost' counter remains as failed upcalls, but
is annotated to make it clear what the meaning is.
6. The enum pmd_stat_type is annotated to make the usage of the
stats counters clear.
7. The subtable lookup stats is renamed to make it clear that it
relates to masked lookups.
8. The PMD stats test is updated to handle the new user stats of
packets received, packets recirculated and average number of datapath
passes per packet.
On top of that introduce a "-pmd <core>" option to the PMD info
commands to filter the output for a single PMD.
Made the pmd-stats-show output a bit more readable by adding a blank
between colon and value.
Signed-off-by: Jan Scheurich <jan.scheurich@ericsson.com>
Co-authored-by: Darrell Ball <dlu998@gmail.com>
Signed-off-by: Darrell Ball <dlu998@gmail.com>
Acked-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Billy O'Mahony <billy.o.mahony@intel.com>
Signed-off: Ian Stokes <ian.stokes@intel.com>
