tailscale/net/art/table.go

163 lines
4.7 KiB
Go
Raw Normal View History

net/art: implement the stride table building block of ART A stride table is an 8-bit routing table implemented as an array binary tree, with a special tree updating function (allot) that enables lightning fast address lookups and reasonably fast insertion and deletion. Insertion, deletion and lookup are all allocation-free. Updates #7781 │ sec/op │ StrideTableInsertion/10/random_order 16.79n ± 2% StrideTableInsertion/10/largest_first 16.83n ± 1% StrideTableInsertion/10/smallest_first 16.83n ± 0% StrideTableInsertion/50/random_order 17.84n ± 1% StrideTableInsertion/50/largest_first 20.04n ± 1% StrideTableInsertion/50/smallest_first 16.39n ± 0% StrideTableInsertion/100/random_order 14.63n ± 0% StrideTableInsertion/100/largest_first 17.45n ± 4% StrideTableInsertion/100/smallest_first 12.98n ± 0% StrideTableInsertion/200/random_order 12.51n ± 4% StrideTableInsertion/200/largest_first 18.36n ± 3% StrideTableInsertion/200/smallest_first 9.609n ± 3% StrideTableDeletion/10/random_order 19.50n ± 1% StrideTableDeletion/10/largest_first 19.34n ± 0% StrideTableDeletion/10/smallest_first 19.43n ± 0% StrideTableDeletion/50/random_order 14.58n ± 1% StrideTableDeletion/50/largest_first 14.27n ± 2% StrideTableDeletion/50/smallest_first 15.51n ± 0% StrideTableDeletion/100/random_order 12.02n ± 3% StrideTableDeletion/100/largest_first 10.64n ± 0% StrideTableDeletion/100/smallest_first 13.21n ± 3% StrideTableDeletion/200/random_order 14.05n ± 4% StrideTableDeletion/200/largest_first 9.288n ± 5% StrideTableDeletion/200/smallest_first 18.51n ± 1% StrideTableGet 0.5010n ± 0% │ routes/s │ StrideTableInsertion/10/random_order 59.55M ± 2% StrideTableInsertion/10/largest_first 59.42M ± 1% StrideTableInsertion/10/smallest_first 59.43M ± 0% StrideTableInsertion/50/random_order 56.04M ± 1% StrideTableInsertion/50/largest_first 49.91M ± 1% StrideTableInsertion/50/smallest_first 61.00M ± 0% StrideTableInsertion/100/random_order 68.35M ± 0% StrideTableInsertion/100/largest_first 57.32M ± 3% StrideTableInsertion/100/smallest_first 77.06M ± 0% StrideTableInsertion/200/random_order 79.93M ± 4% StrideTableInsertion/200/largest_first 54.47M ± 3% StrideTableInsertion/200/smallest_first 104.1M ± 3% StrideTableDeletion/10/random_order 51.28M ± 1% StrideTableDeletion/10/largest_first 51.70M ± 0% StrideTableDeletion/10/smallest_first 51.48M ± 0% StrideTableDeletion/50/random_order 68.60M ± 1% StrideTableDeletion/50/largest_first 70.09M ± 2% StrideTableDeletion/50/smallest_first 64.45M ± 0% StrideTableDeletion/100/random_order 83.21M ± 3% StrideTableDeletion/100/largest_first 94.03M ± 0% StrideTableDeletion/100/smallest_first 75.69M ± 3% StrideTableDeletion/200/random_order 71.20M ± 5% StrideTableDeletion/200/largest_first 107.7M ± 5% StrideTableDeletion/200/smallest_first 54.02M ± 1% StrideTableGet 1.996G ± 0% Signed-off-by: David Anderson <danderson@tailscale.com>
2023-04-03 23:29:36 +00:00
// Copyright (c) Tailscale Inc & AUTHORS
// SPDX-License-Identifier: BSD-3-Clause
// Package art provides a routing table that implements the Allotment Routing
// Table (ART) algorithm by Donald Knuth, as described in the paper by Yoichi
// Hariguchi.
//
// ART outperforms the traditional radix tree implementations for route lookups,
// insertions, and deletions.
//
// For more information, see Yoichi Hariguchi's paper:
// https://cseweb.ucsd.edu//~varghese/TEACH/cs228/artlookup.pdf
package art
net/art: implement the Table type, a multi-level art route table. Updates #7781 │ sec/op │ TableInsertion/ipv4/10 1.562µ ± 2% TableInsertion/ipv4/100 2.398µ ± 5% TableInsertion/ipv4/1000 2.097µ ± 3% TableInsertion/ipv4/10000 2.756µ ± 4% TableInsertion/ipv4/100000 2.473µ ± 13% TableInsertion/ipv6/10 7.649µ ± 2% TableInsertion/ipv6/100 12.09µ ± 3% TableInsertion/ipv6/1000 14.84µ ± 5% TableInsertion/ipv6/10000 14.72µ ± 8% TableInsertion/ipv6/100000 13.23µ ± 41% TableDelete/ipv4/10 378.4n ± 5% TableDelete/ipv4/100 366.9n ± 3% TableDelete/ipv4/1000 418.6n ± 3% TableDelete/ipv4/10000 609.2n ± 11% TableDelete/ipv4/100000 679.2n ± 28% TableDelete/ipv6/10 504.2n ± 4% TableDelete/ipv6/100 959.5n ± 12% TableDelete/ipv6/1000 1.436µ ± 6% TableDelete/ipv6/10000 1.772µ ± 15% TableDelete/ipv6/100000 1.172µ ± 113% TableGet/ipv4/10 32.14n ± 11% TableGet/ipv4/100 38.58n ± 2% TableGet/ipv4/1000 45.03n ± 2% TableGet/ipv4/10000 52.90n ± 7% TableGet/ipv4/100000 135.2n ± 11% TableGet/ipv6/10 41.55n ± 1% TableGet/ipv6/100 44.78n ± 2% TableGet/ipv6/1000 49.03n ± 2% TableGet/ipv6/10000 65.38n ± 5% TableGet/ipv6/100000 525.0n ± 39% │ avg-B/op │ TableInsertion/ipv4/10 25.18Ki ± 0% TableInsertion/ipv4/100 17.63Ki ± 0% TableInsertion/ipv4/1000 14.14Ki ± 0% TableInsertion/ipv4/10000 12.92Ki ± 0% TableInsertion/ipv4/100000 11.13Ki ± 0% TableInsertion/ipv6/10 76.87Ki ± 0% TableInsertion/ipv6/100 98.33Ki ± 0% TableInsertion/ipv6/1000 91.44Ki ± 0% TableInsertion/ipv6/10000 90.39Ki ± 0% TableInsertion/ipv6/100000 87.19Ki ± 0% TableDelete/ipv4/10 3.230 ± 0% TableDelete/ipv4/100 4.020 ± 0% TableDelete/ipv4/1000 3.990 ± 0% TableDelete/ipv4/10000 4.000 ± 0% TableDelete/ipv4/100000 4.000 ± 0% TableDelete/ipv6/10 16.00 ± 0% TableDelete/ipv6/100 16.00 ± 0% TableDelete/ipv6/1000 16.00 ± 0% TableDelete/ipv6/10000 16.00 ± 0% TableDelete/ipv6/100000 16.00 ± 0% │ avg-allocs/op │ TableInsertion/ipv4/10 2.900 ± 0% TableInsertion/ipv4/100 2.330 ± 0% TableInsertion/ipv4/1000 2.070 ± 0% TableInsertion/ipv4/10000 1.980 ± 0% TableInsertion/ipv4/100000 1.840 ± 0% TableInsertion/ipv6/10 6.800 ± 0% TableInsertion/ipv6/100 8.420 ± 0% TableInsertion/ipv6/1000 7.900 ± 0% TableInsertion/ipv6/10000 7.820 ± 0% TableInsertion/ipv6/100000 7.580 ± 0% TableDelete/ipv4/10 1.000 ± 0% TableDelete/ipv4/100 1.000 ± 0% TableDelete/ipv4/1000 1.000 ± 0% TableDelete/ipv4/10000 1.000 ± 0% TableDelete/ipv4/100000 1.000 ± 0% TableDelete/ipv6/10 1.000 ± 0% TableDelete/ipv6/100 1.000 ± 0% TableDelete/ipv6/1000 1.000 ± 0% TableDelete/ipv6/10000 1.000 ± 0% TableDelete/ipv6/100000 1.000 ± 0% │ routes/s │ TableInsertion/ipv4/10 640.3k ± 2% TableInsertion/ipv4/100 417.1k ± 5% TableInsertion/ipv4/1000 477.0k ± 3% TableInsertion/ipv4/10000 362.8k ± 5% TableInsertion/ipv4/100000 404.5k ± 15% TableInsertion/ipv6/10 130.7k ± 1% TableInsertion/ipv6/100 82.69k ± 3% TableInsertion/ipv6/1000 67.37k ± 5% TableInsertion/ipv6/10000 67.93k ± 9% TableInsertion/ipv6/100000 75.63k ± 29% TableDelete/ipv4/10 2.642M ± 6% TableDelete/ipv4/100 2.726M ± 3% TableDelete/ipv4/1000 2.389M ± 3% TableDelete/ipv4/10000 1.641M ± 12% TableDelete/ipv4/100000 1.472M ± 27% TableDelete/ipv6/10 1.984M ± 4% TableDelete/ipv6/100 1.042M ± 11% TableDelete/ipv6/1000 696.5k ± 6% TableDelete/ipv6/10000 564.4k ± 13% TableDelete/ipv6/100000 853.6k ± 53% │ addrs/s │ TableGet/ipv4/10 31.11M ± 10% TableGet/ipv4/100 25.92M ± 2% TableGet/ipv4/1000 22.21M ± 2% TableGet/ipv4/10000 18.91M ± 8% TableGet/ipv4/100000 7.397M ± 12% TableGet/ipv6/10 24.07M ± 1% TableGet/ipv6/100 22.33M ± 2% TableGet/ipv6/1000 20.40M ± 2% TableGet/ipv6/10000 15.30M ± 5% TableGet/ipv6/100000 1.905M ± 28% │ B/op │ TableGet/ipv4/10 4.000 ± 0% TableGet/ipv4/100 4.000 ± 0% TableGet/ipv4/1000 4.000 ± 0% TableGet/ipv4/10000 4.000 ± 0% TableGet/ipv4/100000 4.000 ± 0% TableGet/ipv6/10 16.00 ± 0% TableGet/ipv6/100 16.00 ± 0% TableGet/ipv6/1000 16.00 ± 0% TableGet/ipv6/10000 16.00 ± 0% TableGet/ipv6/100000 16.00 ± 0% │ allocs/op │ TableGet/ipv4/10 1.000 ± 0% TableGet/ipv4/100 1.000 ± 0% TableGet/ipv4/1000 1.000 ± 0% TableGet/ipv4/10000 1.000 ± 0% TableGet/ipv4/100000 1.000 ± 0% TableGet/ipv6/10 1.000 ± 0% TableGet/ipv6/100 1.000 ± 0% TableGet/ipv6/1000 1.000 ± 0% TableGet/ipv6/10000 1.000 ± 0% TableGet/ipv6/100000 1.000 ± 0% Signed-off-by: David Anderson <danderson@tailscale.com>
2023-04-04 16:00:51 +00:00
import (
"bytes"
"fmt"
"io"
"net/netip"
"strings"
)
// Table is an IPv4 and IPv6 routing table.
type Table[T any] struct {
v4 strideTable[T]
v6 strideTable[T]
}
// Get does a route lookup for addr and returns the associated value, or nil if
// no route matched.
func (t *Table[T]) Get(addr netip.Addr) *T {
st := &t.v4
if addr.Is6() {
st = &t.v6
}
var ret *T
for _, stride := range addr.AsSlice() {
rt, child := st.getValAndChild(stride)
if rt != nil {
// Found a more specific route than whatever we found previously,
// keep a note.
ret = rt
}
if child == nil {
// No sub-routes further down, whatever we have recorded in ret is
// the result.
return ret
}
st = child
}
// Unreachable because Insert/Delete won't allow the leaf strideTables to
// have children, so we must return via the nil check in the loop.
panic("unreachable")
}
// Insert adds pfx to the table, with value val.
// If pfx is already present in the table, its value is set to val.
func (t *Table[T]) Insert(pfx netip.Prefix, val *T) {
if val == nil {
panic("Table.Insert called with nil value")
}
st := &t.v4
if pfx.Addr().Is6() {
st = &t.v6
}
bs := pfx.Addr().AsSlice()
i := 0
numBits := pfx.Bits()
// The strideTable we want to insert into is potentially at the end of a
// chain of parent tables, each one encoding successive 8 bits of the
// prefix. Navigate downwards, allocating child tables as needed, until we
// find the one this prefix belongs in.
for numBits > 8 {
st = st.getOrCreateChild(bs[i])
i++
numBits -= 8
}
// Finally, insert the remaining 0-8 bits of the prefix into the child
// table.
st.insert(bs[i], numBits, val)
}
// Delete removes pfx from the table, if it is present.
func (t *Table[T]) Delete(pfx netip.Prefix) {
st := &t.v4
if pfx.Addr().Is6() {
st = &t.v6
}
bs := pfx.Addr().AsSlice()
i := 0
numBits := pfx.Bits()
// Deletion may drive the refcount of some strideTables down to zero. We
// need to clean up these dangling tables, so we have to keep track of which
// tables we touch on the way down, and which strideEntry index each child
// is registered in.
strideTables := [16]*strideTable[T]{st}
var strideIndexes [16]int
// Similar to Insert, navigate down the tree of strideTables, looking for
// the one that houses the last 0-8 bits of the prefix to delete.
//
// The only difference is that here, we don't create missing child tables.
// If a child necessary to pfx is missing, then the pfx cannot exist in the
// Table, and we can exit early.
for numBits > 8 {
child, idx := st.getChild(bs[i])
if child == nil {
// Prefix can't exist in the table, one of the necessary
// strideTables doesn't exit.
return
}
// Note that the strideIndex and strideTables entries are off-by-one.
// The child table pointer is recorded at i+1, but it is referenced by a
// particular index in the parent table, at index i.
strideIndexes[i] = idx
i++
strideTables[i] = child
numBits -= 8
st = child
}
if st.delete(bs[i], numBits) == nil {
// Prefix didn't exist in the expected strideTable, refcount hasn't
// changed, no need to run through cleanup.
return
}
// st.delete reduced st's refcount by one, so we may be hanging onto a chain
// of redundant strideTables. Walk back up the path we recorded in the
// descent loop, deleting tables until we encounter one that still has other
// refs (or we hit the root strideTable, which is never deleted).
for i > 0 && strideTables[i].refs == 0 {
strideTables[i-1].deleteChild(strideIndexes[i-1])
i--
}
}
// debugSummary prints the tree of allocated strideTables in t, with each
// strideTable's refcount.
func (t *Table[T]) debugSummary() string {
var ret bytes.Buffer
fmt.Fprintf(&ret, "v4: ")
strideSummary(&ret, &t.v4, 0)
fmt.Fprintf(&ret, "v6: ")
strideSummary(&ret, &t.v6, 0)
return ret.String()
}
func strideSummary[T any](w io.Writer, st *strideTable[T], indent int) {
fmt.Fprintf(w, "%d refs\n", st.refs)
indent += 2
for i := firstHostIndex; i <= lastHostIndex; i++ {
if child := st.entries[i].child; child != nil {
addr, len := inversePrefixIndex(i)
fmt.Fprintf(w, "%s%d/%d: ", strings.Repeat(" ", indent), addr, len)
strideSummary(w, child, indent)
}
}
}