New cache implementation and prefetch handing in mw/cache (#731)

* cache: add sharded cache implementation

Add Cache impl and a few tests. This cache is 256-way sharded, mainly
so each shard has it's own lock. The main cache structure is a readonly
jump plane into the right shard.

This should remove the single lock contention on the main lock and
provide more concurrent throughput - Obviously this hasn't been tested
or measured.

The key into the cache was made a uint32 (hash.fnv) and the hashing op
is not using strings.ToLower anymore remove any GC in that code path.

* here too

* Minimum shard size

* typos

* blurp

* small cleanups no defer

* typo

* Add freq based on Johns idea

* cherry-pick conflict resolv

* typo

* update from early code review from john

* add prefetch to the cache

* mw/cache: add prefetch

* remove println

* remove comment

* Fix tests

* Test prefetch in setup

* Add start of cache

* try add diff cache options

* Add hacky testcase

* not needed

* allow the use of a percentage for prefetch

If the TTL falls below xx% do a prefetch, if the record was popular.
Some other fixes and correctly prefetch only popular records.

1 files changed, 5 insertions, 6 deletions

diff --git a/middleware/dnssec/cache.go b/middleware/dnssec/cache.go
index 2153c84cb..ea95b73b4 100644
--- a/middleware/dnssec/cache.go
+++ b/middleware/dnssec/cache.go
@@ -2,14 +2,13 @@ package dnssec
 
 import (
 	"hash/fnv"
-	"strconv"
 
 	"github.com/miekg/dns"
 )
 
-// Key serializes the RRset and return a signature cache key.
-func key(rrs []dns.RR) string {
-	h := fnv.New64()
+// hash serializes the RRset and return a signature cache key.
+func hash(rrs []dns.RR) uint32 {
+	h := fnv.New32()
 	buf := make([]byte, 256)
 	for _, r := range rrs {
 		off, err := dns.PackRR(r, buf, 0, nil, false)
@@ -18,6 +17,6 @@ func key(rrs []dns.RR) string {
 		}
 	}
 
-	i := h.Sum64()
-	return strconv.FormatUint(i, 10)
+	i := h.Sum32()
+	return i
 }