Use xxhash64 for bucket-to-stripe distribution

Existing naive prime-based solution was susceptible to scheduling
operations for the subtree of a superbucket in one strand alone,
despite previous attempts to disperse this using prime number
multiplication. This would put a serious limiter on parallelism
for super bucket locality-sensitive reads such as streaming search
visitors.

2 files changed, 10 insertions, 6 deletions

diff --git a/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.cpp b/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.cpp
index 5627ade2a11..fbdbac27b7c 100644
--- a/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.cpp
+++ b/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.cpp
@@ -14,6 +14,10 @@
 #include <vespa/storageapi/message/stat.h>
 #include <vespa/vespalib/stllike/hash_map.hpp>
 #include <vespa/vespalib/util/exceptions.h>
+#ifndef XXH_INLINE_ALL
+#  define XXH_INLINE_ALL // Let XXH64 be inlined for fixed hash size (bucket ID)
+#endif
+#include <xxhash.h>
 
 #include <vespa/log/log.h>
 LOG_SETUP(".persistence.filestor.handler.impl");
@@ -894,6 +898,11 @@ FileStorHandlerImpl::Disk::broadcast()
     }
 }
 
+uint64_t FileStorHandlerImpl::Disk::dispersed_bucket_bits(const document::Bucket& bucket) noexcept {
+    const uint64_t raw_id = bucket.getBucketId().getId();
+    return XXH64(&raw_id, sizeof(uint64_t), 0);
+}
+
 bool
 FileStorHandlerImpl::Disk::schedule(const std::shared_ptr<api::StorageMessage>& msg)
 {
diff --git a/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.h b/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.h
index 7a4f9000e82..00714c291b7 100644
--- a/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.h
+++ b/storage/src/vespa/storage/persistence/filestorage/filestorhandlerimpl.h
@@ -197,12 +197,7 @@ public:
         std::string dumpQueue() const;
         void dumpActiveHtml(std::ostream & os) const;
         void dumpQueueHtml(std::ostream & os) const;
-        static uint64_t dispersed_bucket_bits(const document::Bucket& bucket) noexcept {
-            // Disperse bucket bits by multiplying with the 64-bit FNV-1 prime.
-            // This avoids an inherent affinity between the LSB of a bucket's bits
-            // and the stripe an operation ends up on.
-            return bucket.getBucketId().getId() * 1099511628211ULL;
-        }
+        static uint64_t dispersed_bucket_bits(const document::Bucket& bucket) noexcept;
         // We make a fairly reasonable assumption that there will be less than 64k stripes.
         uint16_t stripe_index(const document::Bucket& bucket) const noexcept {
             return static_cast<uint16_t>(dispersed_bucket_bits(bucket) % _stripes.size());