nft_set_pipapo: Add support for 8-bit lookup groups and dynamic switch
authorStefano Brivio <sbrivio@redhat.com>
Sat, 7 Mar 2020 16:52:33 +0000 (17:52 +0100)
committerPablo Neira Ayuso <pablo@netfilter.org>
Sun, 15 Mar 2020 14:27:43 +0000 (15:27 +0100)
commit4051f43116cdc7dc4df729d4ef80d1ac46a1297f
tree6d7a5e9c32dd691aef8a21e0b33ae60c985b7933
parente807b13cb3e3bcf0f602cb5ef66f7a1988d0e703
nft_set_pipapo: Add support for 8-bit lookup groups and dynamic switch

While grouping matching bits in groups of four saves memory compared
to the more natural choice of 8-bit words (lookup table size is one
eighth), it comes at a performance cost, as the number of lookup
comparisons is doubled, and those also needs bitshifts and masking.

Introduce support for 8-bit lookup groups, together with a mapping
mechanism to dynamically switch, based on defined per-table size
thresholds and hysteresis, between 8-bit and 4-bit groups, as tables
grow and shrink. Empty sets start with 8-bit groups, and per-field
tables are converted to 4-bit groups if they get too big.

An alternative approach would have been to swap per-set lookup
operation functions as needed, but this doesn't allow for different
group sizes in the same set, which looks desirable if some fields
need significantly more matching data compared to others due to
heavier impact of ranges (e.g. a big number of subnets with
relatively simple port specifications).

Allowing different group sizes for the same lookup functions implies
the need for further conditional clauses, whose cost, however,
appears to be negligible in tests.

The matching rate figures below were obtained for x86_64 running
the nft_concat_range.sh "performance" cases, averaged over five
runs, on a single thread of an AMD Epyc 7402 CPU, and for aarch64
on a single thread of a BCM2711 (Raspberry Pi 4 Model B 4GB),
clocked at a stable 2147MHz frequency:

---------------.-----------------------------------.------------.
AMD Epyc 7402  |          baselines, Mpps          | this patch |
 1 thread      |___________________________________|____________|
 3.35GHz       |        |        |        |        |            |
 768KiB L1D$   | netdev |  hash  | rbtree |        |            |
---------------|  hook  |   no   | single | pipapo |   pipapo   |
type   entries |  drop  | ranges | field  | 4 bits | bit switch |
---------------|--------|--------|--------|--------|------------|
net,port       |        |        |        |        |            |
         1000  |   19.0 |   10.4 |    3.8 |    2.8 | 4.0   +43% |
---------------|--------|--------|--------|--------|------------|
port,net       |        |        |        |        |            |
          100  |   18.8 |   10.3 |    5.8 |    5.5 | 6.3   +14% |
---------------|--------|--------|--------|--------|------------|
net6,port      |        |        |        |        |            |
         1000  |   16.4 |    7.6 |    1.8 |    1.3 | 2.1   +61% |
---------------|--------|--------|--------|--------|------------|
port,proto     |        |        |        |        |     [1]    |
        30000  |   19.6 |   11.6 |    3.9 |    0.3 | 0.5   +66% |
---------------|--------|--------|--------|--------|------------|
net6,port,mac  |        |        |        |        |            |
           10  |   16.5 |    5.4 |    4.3 |    2.6 | 3.4   +31% |
---------------|--------|--------|--------|--------|------------|
net6,port,mac, |        |        |        |        |            |
proto    1000  |   16.5 |    5.7 |    1.9 |    1.0 | 1.4   +40% |
---------------|--------|--------|--------|--------|------------|
net,mac        |        |        |        |        |            |
         1000  |   19.0 |    8.4 |    3.9 |    1.7 | 2.5   +47% |
---------------'--------'--------'--------'--------'------------'
[1] Causes switch of lookup table buckets for 'port', not 'proto',
    to 4-bit groups

 ---------------.-----------------------------------.------------.
 BCM2711        |          baselines, Mpps          | this patch |
  1 thread      |___________________________________|____________|
  2147MHz       |        |        |        |        |            |
  32KiB L1D$    | netdev |  hash  | rbtree |        |            |
 ---------------|  hook  |   no   | single | pipapo |   pipapo   |
 type   entries |  drop  | ranges | field  | 4 bits | bit switch |
 ---------------|--------|--------|--------|--------|------------|
 net,port       |        |        |        |        |            |
          1000  |   1.63 |   1.37 |   0.87 |   0.61 | 0.70  +17% |
 ---------------|--------|--------|--------|--------|------------|
 port,net       |        |        |        |        |            |
           100  |   1.64 |   1.36 |   1.02 |   0.78 | 0.81   +4% |
 ---------------|--------|--------|--------|--------|------------|
 net6,port      |        |        |        |        |            |
          1000  |   1.56 |   1.27 |   0.65 |   0.34 | 0.50  +47% |
 ---------------|--------|--------|--------|--------|------------|
 port,proto [2] |        |        |        |        |            |
         10000  |   1.68 |   1.43 |   0.84 |   0.30 | 0.40  +13% |
 ---------------|--------|--------|--------|--------|------------|
 net6,port,mac  |        |        |        |        |            |
            10  |   1.56 |   1.14 |   1.02 |   0.62 | 0.66   +6% |
 ---------------|--------|--------|--------|--------|------------|
 net6,port,mac, |        |        |        |        |            |
 proto    1000  |   1.56 |   1.12 |   0.64 |   0.27 | 0.40  +48% |
 ---------------|--------|--------|--------|--------|------------|
 net,mac        |        |        |        |        |            |
          1000  |   1.63 |   1.26 |   0.87 |   0.41 | 0.53  +29% |
 ---------------'--------'--------'--------'--------'------------'
[2] Using 10000 entries instead of 30000 as it would take way too
    long for the test script to generate all of them

Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
net/netfilter/nft_set_pipapo.c