[ovs-dev] [PATCH 0/5] dpif-netdev: Cuckoo-Distributor implementation

[yipeng1.wang at intel.com]

From: Yipeng Wang <yipeng1.wang at intel.com>

The Datapath Classifier uses tuple space search for flow classification.
The rules are arranged into a set of tuples/subtables (each with a
distinct mask).  Each subtable is implemented as a hash table and lookup
is done with flow keys formed by selecting the bits from the packet header
based on each subtable's mask. Tuple space search will sequentially search
each subtable until a match is found. With a large number of subtables, a
sequential search of the subtables could consume a lot of CPU cycles. In
a testbench with a uniform traffic pattern equally distributed across 20
subtables, we measured that up to 65% of total execution time is attributed
to the megaflow cache lookup.

This patch presents the idea of the two-layer hierarchical lookup, where a
low overhead first level of indirection is accessed first, we call this
level cuckoo distributor (CD). If a flow key has been inserted in the flow
table the first level will indicate with high probability that which
subtable to look into. A lookup is performed on the second level (the
target subtable) to retrieve the result. If the key doesn’t have a match,
then we revert back to the sequential search of subtables. The patch is
partially inspired by earlier concepts proposed in "simTable"[1] and
"Cuckoo Filter"[2], and DPDK's Cuckoo Hash implementation.

This patch can improve the already existing Subtable Ranking when traffic
data has high entropy. Subtable Ranking helps minimize the number of
traversed subtables when most of the traffic hit the same subtable.
However, in the case of high entropy traffic such as traffic coming from
a physical port, multiple subtables could be hit with a similar frequency.
In this case the average subtable lookups per hit would be much greater
than 1. In addition, CD can adaptively turn off when it finds the traffic
mostly hit one subtable. Thus, CD will not be an overhead when Subtable
Ranking works well.

Scheme:

 -------
|  CD   |
 -------
       \
        \
 -----  -----     -----
|sub  ||sub  |...|sub  |
|table||table|   |table|
 -----  -----     -----

Evaluation:

We create set of rules with various src IP. We feed traffic containing various
numbers of flows with various src IP and dst IP. All the flows hit 10/20/30
rules creating 10/20/30 subtables.

The table below shows the preliminary continuous testing results (full line
speed test) we collected with a uni-directional phy-to-phy setup. The
machine we tested on is a Xeon E5 server running with 2.2GHz cores. OvS
runs with 1 PMD. We use Spirent as the hardware traffic generator.

AVX2 data:
20k flows:
no.subtable: 10          20          30
cd-ovs       4267332     3478251     3126763
orig-ovs     3260883     2174551     1689981
speedup      1.31x       1.60x       1.85x

100k flows:
no.subtable: 10          20          30
cd-ovs       4015783     3276100     2970645
orig-ovs     2692882     1711955     1302321
speedup      1.49x       1.91x       2.28x

1M flows:
no.subtable: 10          20          30
cd-ovs       3895961     3170530     2968555
orig-ovs     2683455     1646227     1240501
speedup      1.45x       1.92x       2.39x

Scalar data:
1M flows:
no.subtable: 10          20          30
cd-ovs       3658328     3028111     2863329
orig_ovs     2683455     1646227     1240501
speedup      1.36x       1.84x       2.31x

[1] H. Lee and B. Lee, Approaches for improving tuple space search-based
table lookup, ICTC '15
[2] B. Fan, D. G. Andersen, M. Kaminsky, and M. D. Mitzenmacher,
Cuckoo Filter: Practically Better Than Bloom, CoNEXT '14

This patch set is created based on commit
a13784ba95efeb5a1f77253df40d433a1ce60087

The previous RFC on mailing list are at:
https://mail.openvswitch.org/pipermail/ovs-dev/2017-May/331834.html
https://mail.openvswitch.org/pipermail/ovs-dev/2017-April/330570.html

Signed-off-by: Yipeng Wang <yipeng1.wang at intel.com>
Signed-off-by: Charlie Tai <charlie.tai at intel.com>
Co-authored-by: Charlie Tai <charlie.tai at intel.com>
Signed-off-by: Sameh Gobriel <sameh.gobriel at intel.com>
Co-authored-by: Sameh Gobriel <sameh.gobriel at intel.com>
Signed-off-by: Ren Wang <ren.wang at intel.com>
Co-authored-by: Ren Wang <ren.wang at intel.com>
Signed-off-by: Antonio Fischetti <antonio.fischetti at intel.com>
Co-authored-by: Antonio Fischetti <antonio.fischetti at intel.com>


Yipeng Wang (5):
  dpif-netdev: Basic CD feature with scalar lookup.
  dpif-netdev: Add AVX2 implementation for CD lookup.
  dpif-netdev: Add CD statistics
  dpif-netdev: Add adaptive CD mechanism
  unit-test: Add a delay for CD initialization.

 lib/dpif-netdev.c     | 566 +++++++++++++++++++++++++++++++++++++++++++++++++-
 tests/ofproto-dpif.at |   3 +
 2 files changed, 558 insertions(+), 11 deletions(-)

-- 
1.9.1