The bart package provides some Balanced Routing Tables (BART) for fastest IP-to-CIDR lookups and related tasks such as:

• ACL determine extremely fast whether an IP address matches any of millions of CIDR rules.
• RIB handle very large routing tables with low memory overhead, while keeping lookups fast.
• FIB high-speed lookups, achieve LPM in constant-time for packet forwarding in the datapath.

BART is designed for workloads where both speed and/or memory efficiency matter, making it a best fit for firewalls, routers, or any system that needs large-scale IP prefix matching.

BART is implemented as a multibit trie with a fixed stride of 8 bits, using a fast mapping function derived from Donald E. Knuth’s Allotment Routing Table (ART) algorithm, to map the possible prefixes at each level into a complete binary tree.

BART implements three different routing tables, each optimized for specific use cases:

• bart.Lite
• bart.Table
• bart.Fast

For bart.Table this binary tree is represented with popcount‑compressed sparse arrays for level compression. Combined with a novel path and fringe compression, this design reduces memory consumption by nearly two orders of magnitude compared to classical ART.

For bart.Fast this binary tree is represented with fixed arrays without level compression (classical ART), but combined with the same novel path and fringe compression from BART. This design reduces memory consumption by more than an order of magnitude compared to classical ART and thus makes ART usable in the first place for large routing tables.

bart.Lite is a special form of bart.Table, but without a payload, and therefore has the lowest memory overhead while maintaining the same lookup times.

package main

import (
  "net/netip"

  "github.com/gaissmai/bart"
)

var a = netip.MustParseAddr
var p = netip.MustParsePrefix

func main() {
  allowList := new(bart.Lite)

  // Add allowed networks
  allowList.Insert(p("192.168.0.0/16"))
  allowList.Insert(p("2001:db8::/32"))
  allowList.Insert(p("10.0.0.0/24"))

  // Test some IPs with Contains (longest-prefix match)
  testIPs := []netip.Addr{
    a("192.168.1.100"), // allowed (matches 192.168.0.0/16)
    a("2001:db8::1"),   // allowed (matches 2001:db8::/32)
    a("172.16.0.1"),    // denied (no match)
  }

  for _, ip := range testIPs {
    if allowList.Contains(ip) {
      fmt.Printf("✅ %s is allowed\n", ip)
    } else {
      fmt.Printf("❌ %s is denied\n", ip)
    }
  }

  // Exact-prefix checks with Get (not longest-prefix match)
  exactPrefixes := []netip.Prefix{
    p("192.168.0.0/16"),  // exists exactly
    p("192.168.1.0/24"),  // doesn't exist (would be covered by /16, but not stored exactly)
    p("10.0.0.0/24"),     // exists exactly
    p("10.0.0.0/16"),     // doesn't exist (only /24 is stored)
  }

  fmt.Println("\nExact-prefix checks:")
  for _, pfx := range exactPrefixes {
    if allowList.Get(pfx) {
      fmt.Printf("✅ Exact prefix %s exists in table\n", pfx)
    } else {
      fmt.Printf("❌ Exact prefix %s not found in table\n", pfx)
    }
  }
}

A more detailed description can be found in NODETYPES.md.

• Recommended for most routing table use cases
• Near-optimal per-level performance with excellent memory efficiency
• Perfect balance for both IPv4 and IPv6 routing tables (use it for RIB)

• Specialized for prefix-only operations, no payload
• Same per-level performance as bart.Table[V] but 35% less memory
• Ideal for IPv4/IPv6 allowlists and set-based operations (use it for ACL)

• 50-100% faster when memory constraints allow
• Best choice for lookup-intensive applications (use it for FIB)

The BART algorithm is based on fixed-size bit vectors and precomputed lookup tables. Lookups are executed entirely with fast, cache-resident bitmask operations, which modern CPUs accelerate using specialized instructions such as POPCNT, LZCNT, and TZCNT.

For maximum performance, specify the CPU feature set when compiling. See the Go minimum requirements for details.

# On ARM64, Go auto-selects CPU instructions.
# Example for AMD64, choose v2/v3/v4 to match your CPU features.
GOAMD64=v3 go build

Critical loops over these fixed-size bitsets can be unrolled for additional speed, ensuring predictable memory access and efficient use of CPU pipelines.

func (b *BitSet256) popcnt() (cnt int) {
  cnt += bits.OnesCount64(b[0])
  cnt += bits.OnesCount64(b[1])
  cnt += bits.OnesCount64(b[2])
  cnt += bits.OnesCount64(b[3])
  return
}

Future Go versions with SIMD intrinsics for uint64 vectors may unlock additional speedups on compatible hardware.

There are examples demonstrating how to use bart concurrently with multiple readers and writers. Readers can always access the table lock‑free. Writers synchronize with a mutex so that only one writer modifies the persistent table at a time, without relying on CAS, which can be problematic with multiple long‑running writers.

The combination of lock-free concurrency, fast lookup and update times and low memory consumption provides clear advantages for any routing daemon.

But as always, it depends on the specific use case.

See the concurrent tests for concrete examples of this pattern:

• ExampleFast
• ExampleLite
• ExampleTable

Beyond high-performance prefix matching, BART also excels at detecting overlaps between two routing tables. In internal benchmarks the check runs in a few nanoseconds per query with zero heap allocations on a modern CPU.

BART has a rich API for CRUD, lookup, comparison, iteration, serialization and persistence.

Table and Fast expose the identical API, while Lite deviates in its methods from the common API when it comes to the payload, since Lite has no payload.

import "github.com/gaissmai/bart"

type Table[V any] struct {
  // Has unexported fields.
}

func (t *Table[V]) Contains(netip.Addr) bool
func (t *Table[V]) Lookup(netip.Addr) (V, bool)

func (t *Table[V]) LookupPrefix(netip.Prefix) (V, bool)
func (t *Table[V]) LookupPrefixLPM(netip.Prefix) (netip.Prefix, V, bool)

func (t *Table[V]) Get(netip.Prefix) (V, bool)
func (t *Table[V]) Insert(netip.Prefix, V)
func (t *Table[V]) Delete(netip.Prefix)
func (t *Table[V]) Modify(netip.Prefix, cb func(V, bool) (V, bool))

func (t *Table[V]) InsertPersist(netip.Prefix, V) *Table[V]
func (t *Table[V]) DeletePersist(netip.Prefix) *Table[V]
func (t *Table[V]) ModifyPersist(netip.Prefix, cb func(V, bool) (V, bool)) *Table[V]

func (t *Table[V]) Clone() *Table[V]
func (t *Table[V]) Union(o *Table[V])
func (t *Table[V]) UnionPersist(o *Table[V]) *Table[V]

func (t *Table[V]) OverlapsPrefix(netip.Prefix) bool

func (t *Table[V]) Overlaps(o *Table[V]) bool
func (t *Table[V]) Overlaps4(o *Table[V]) bool
func (t *Table[V]) Overlaps6(o *Table[V]) bool

func (t *Table[V]) Equal(o *Table[V]) bool

func (t *Table[V]) Subnets(netip.Prefix) iter.Seq2[netip.Prefix, V]
func (t *Table[V]) Supernets(netip.Prefix) iter.Seq2[netip.Prefix, V]

func (t *Table[V]) All() iter.Seq2[netip.Prefix, V]
func (t *Table[V]) All4() iter.Seq2[netip.Prefix, V]
func (t *Table[V]) All6() iter.Seq2[netip.Prefix, V]

func (t *Table[V]) AllSorted() iter.Seq2[netip.Prefix, V]
func (t *Table[V]) AllSorted4() iter.Seq2[netip.Prefix, V]
func (t *Table[V]) AllSorted6() iter.Seq2[netip.Prefix, V]

func (t *Table[V]) Size() int
func (t *Table[V]) Size4() int
func (t *Table[V]) Size6() int

func (t *Table[V]) Fprint(w io.Writer) error
func (t *Table[V]) MarshalText() ([]byte, error)
func (t *Table[V]) MarshalJSON() ([]byte, error)

func (t *Table[V]) DumpList4() []DumpListNode[V]
func (t *Table[V]) DumpList6() []DumpListNode[V]

Please see the extensive benchmarks comparing bart with other IP routing table implementations.

Just a teaser, Fast.Lookup against the Tier1 full Internet routing table with random IP address probes:

$ GOAMD64=v3 go test -run=xxx -bench=FullFastM/Lookup$ -cpu=1
goos: linux
goarch: amd64
pkg: github.com/gaissmai/bart
cpu: Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz
BenchmarkFullFastMatch4/Lookup           124476082          9.63 ns/op
BenchmarkFullFastMatch6/Lookup           91032597          13.20 ns/op
BenchmarkFullFastMiss4/Lookup            134171480          8.93 ns/op
BenchmarkFullFastMiss6/Lookup            75634396          15.79 ns/op

The package is currently released as a pre-v1 version, which gives the author the freedom to break backward compatibility to help improve the API as he learns which initial design decisions would need to be revisited to better support the use cases that the library solves for.

These occurrences are expected to be rare in frequency and the API is already quite stable.

Please open an issue for discussion before sending a pull request.

Standing on the shoulders of giants.

Credits for many inspirations go to

• the clever folks at Tailscale,
• to Daniel Lemire for his inspiring blog,
• to Donald E. Knuth for the Allotment Routing Table ART algorithm and
• to Yoichi Hariguchi who deciphered it for us mere mortals

And last but not least to the Go team who do a wonderful job!

MIT