crypto, pow, vendor: hash optimizations, mmap ethash

3 files changed, 403 insertions, 167 deletions

diff --git a/pow/ethash.go b/pow/ethash.go
index 0af1904b6..dbe8ff077 100644
--- a/pow/ethash.go
+++ b/pow/ethash.go
@@ -17,20 +17,21 @@
 package pow
 
 import (
-	"bufio"
 	"bytes"
 	"errors"
 	"fmt"
-	"io/ioutil"
 	"math"
 	"math/big"
 	"math/rand"
 	"os"
 	"path/filepath"
+	"reflect"
+	"strconv"
 	"sync"
 	"time"
+	"unsafe"
 
-	"github.com/ethereum/go-ethereum/common"
+	mmap "github.com/edsrzf/mmap-go"
 	"github.com/ethereum/go-ethereum/common/hexutil"
 	"github.com/ethereum/go-ethereum/log"
 	metrics "github.com/rcrowley/go-metrics"
@@ -57,10 +58,89 @@ var (
 	dumpMagic = hexutil.MustDecode("0xfee1deadbaddcafe")
 )
 
+// isLittleEndian returns whether the local system is running in little or big
+// endian byte order.
+func isLittleEndian() bool {
+	n := uint32(0x01020304)
+	return *(*byte)(unsafe.Pointer(&n)) == 0x04
+}
+
+// memoryMap tries to memory map a file of uint32s for read only access.
+func memoryMap(path string) (*os.File, mmap.MMap, []uint32, error) {
+	file, err := os.OpenFile(path, os.O_RDONLY, 0644)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+	mem, buffer, err := memoryMapFile(file, false)
+	if err != nil {
+		file.Close()
+		return nil, nil, nil, err
+	}
+	return file, mem, buffer, err
+}
+
+// memoryMapFile tries to memory map an already opened file descriptor.
+func memoryMapFile(file *os.File, write bool) (mmap.MMap, []uint32, error) {
+	// Try to memory map the file
+	flag := mmap.RDONLY
+	if write {
+		flag = mmap.RDWR
+	}
+	mem, err := mmap.Map(file, flag, 0)
+	if err != nil {
+		return nil, nil, err
+	}
+	// Yay, we managed to memory map the file, here be dragons
+	header := *(*reflect.SliceHeader)(unsafe.Pointer(&mem))
+	header.Len /= 4
+	header.Cap /= 4
+
+	return mem, *(*[]uint32)(unsafe.Pointer(&header)), nil
+}
+
+// memoryMapAndGenerate tries to memory map a temporary file of uint32s for write
+// access, fill it with the data from a generator and then move it into the final
+// path requested.
+func memoryMapAndGenerate(path string, size uint64, generator func(buffer []uint32)) (*os.File, mmap.MMap, []uint32, error) {
+	// Ensure the data folder exists
+	if err := os.MkdirAll(filepath.Dir(path), 0755); err != nil {
+		return nil, nil, nil, err
+	}
+	// Create a huge temporary empty file to fill with data
+	temp := path + "." + strconv.Itoa(rand.Int())
+
+	dump, err := os.Create(temp)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+	if err = dump.Truncate(int64(size)); err != nil {
+		return nil, nil, nil, err
+	}
+	// Memory map the file for writing and fill it with the generator
+	mem, buffer, err := memoryMapFile(dump, true)
+	if err != nil {
+		dump.Close()
+		return nil, nil, nil, err
+	}
+	generator(buffer)
+
+	if err := mem.Flush(); err != nil {
+		mem.Unmap()
+		dump.Close()
+		return nil, nil, nil, err
+	}
+	os.Rename(temp, path)
+	return dump, mem, buffer, nil
+}
+
 // cache wraps an ethash cache with some metadata to allow easier concurrent use.
 type cache struct {
-	epoch uint64     // Epoch for which this cache is relevant
-	cache []uint32   // The actual cache data content
+	epoch uint64 // Epoch for which this cache is relevant
+
+	dump *os.File  // File descriptor of the memory mapped cache
+	mmap mmap.MMap // Memory map itself to unmap before releasing
+
+	cache []uint32   // The actual cache data content (may be memory mapped)
 	used  time.Time  // Timestamp of the last use for smarter eviction
 	once  sync.Once  // Ensures the cache is generated only once
 	lock  sync.Mutex // Ensures thread safety for updating the usage time
@@ -71,57 +151,72 @@ func (c *cache) generate(dir string, limit int, test bool) {
 	c.once.Do(func() {
 		// If we have a testing cache, generate and return
 		if test {
-			rawCache := generateCache(1024, seedHash(c.epoch*epochLength+1))
-			c.cache = prepare(1024, bytes.NewReader(rawCache))
+			c.cache = make([]uint32, 1024/4)
+			generateCache(c.cache, c.epoch, seedHash(c.epoch*epochLength+1))
 			return
 		}
-		// Full cache generation is needed, check cache dir for existing data
+		// If we don't store anything on disk, generate and return
 		size := cacheSize(c.epoch*epochLength + 1)
 		seed := seedHash(c.epoch*epochLength + 1)
 
-		path := filepath.Join(dir, fmt.Sprintf("cache-R%d-%x", algorithmRevision, seed))
-		logger := log.New("seed", hexutil.Bytes(seed))
+		if dir == "" {
+			c.cache = make([]uint32, size/4)
+			generateCache(c.cache, c.epoch, seed)
+			return
+		}
+		// Disk storage is needed, this will get fancy
+		endian := "le"
+		if !isLittleEndian() {
+			endian = "be"
+		}
+		path := filepath.Join(dir, fmt.Sprintf("cache-R%d-%x.%s", algorithmRevision, seed, endian))
+		logger := log.New("epoch", c.epoch)
+
+		// Try to load the file from disk and memory map it
+		var err error
+		c.dump, c.mmap, c.cache, err = memoryMap(path)
+		if err == nil {
+			logger.Debug("Loaded old ethash cache from disk")
+			return
+		}
+		logger.Debug("Failed to load old ethash cache", "err", err)
 
-		if dir != "" {
-			dump, err := os.Open(path)
-			if err == nil {
-				logger.Info("Loading ethash cache from disk")
-				start := time.Now()
-				c.cache = prepare(size, bufio.NewReader(dump))
-				logger.Info("Loaded ethash cache from disk", "elapsed", common.PrettyDuration(time.Since(start)))
+		// No previous cache available, create a new cache file to fill
+		c.dump, c.mmap, c.cache, err = memoryMapAndGenerate(path, size, func(buffer []uint32) { generateCache(buffer, c.epoch, seed) })
+		if err != nil {
+			logger.Error("Failed to generate mapped ethash cache", "err", err)
 
-				dump.Close()
-				return
-			}
+			c.cache = make([]uint32, size/4)
+			generateCache(c.cache, c.epoch, seed)
 		}
-		// No previous disk cache was available, generate on the fly
-		rawCache := generateCache(size, seed)
-		c.cache = prepare(size, bytes.NewReader(rawCache))
-
-		// If a cache directory is given, attempt to serialize for next time
-		if dir != "" {
-			// Store the ethash cache to disk
-			start := time.Now()
-			if err := os.MkdirAll(filepath.Dir(path), os.ModePerm); err != nil {
-				logger.Error("Failed to create ethash cache dir", "err", err)
-			} else if err := ioutil.WriteFile(path, rawCache, os.ModePerm); err != nil {
-				logger.Error("Failed to write ethash cache to disk", "err", err)
-			} else {
-				logger.Info("Stored ethash cache to disk", "elapsed", common.PrettyDuration(time.Since(start)))
-			}
-			// Iterate over all previous instances and delete old ones
-			for ep := int(c.epoch) - limit; ep >= 0; ep-- {
-				seed := seedHash(uint64(ep)*epochLength + 1)
-				path := filepath.Join(dir, fmt.Sprintf("cache-R%d-%x", algorithmRevision, seed))
-				os.Remove(path)
-			}
+		// Iterate over all previous instances and delete old ones
+		for ep := int(c.epoch) - limit; ep >= 0; ep-- {
+			seed := seedHash(uint64(ep)*epochLength + 1)
+			path := filepath.Join(dir, fmt.Sprintf("cache-R%d-%x.%s", algorithmRevision, seed, endian))
+			os.Remove(path)
 		}
 	})
 }
 
+// release closes any file handlers and memory maps open.
+func (c *cache) release() {
+	if c.mmap != nil {
+		c.mmap.Unmap()
+		c.mmap = nil
+	}
+	if c.dump != nil {
+		c.dump.Close()
+		c.dump = nil
+	}
+}
+
 // dataset wraps an ethash dataset with some metadata to allow easier concurrent use.
 type dataset struct {
-	epoch   uint64     // Epoch for which this cache is relevant
+	epoch uint64 // Epoch for which this cache is relevant
+
+	dump *os.File  // File descriptor of the memory mapped cache
+	mmap mmap.MMap // Memory map itself to unmap before releasing
+
 	dataset []uint32   // The actual cache data content
 	used    time.Time  // Timestamp of the last use for smarter eviction
 	once    sync.Once  // Ensures the cache is generated only once
@@ -129,78 +224,91 @@ type dataset struct {
 }
 
 // generate ensures that the dataset content is generated before use.
-func (d *dataset) generate(dir string, limit int, test bool, discard bool) {
+func (d *dataset) generate(dir string, limit int, test bool) {
 	d.once.Do(func() {
 		// If we have a testing dataset, generate and return
 		if test {
-			rawCache := generateCache(1024, seedHash(d.epoch*epochLength+1))
-			intCache := prepare(1024, bytes.NewReader(rawCache))
+			cache := make([]uint32, 1024/4)
+			generateCache(cache, d.epoch, seedHash(d.epoch*epochLength+1))
 
-			rawDataset := generateDataset(32*1024, intCache)
-			d.dataset = prepare(32*1024, bytes.NewReader(rawDataset))
+			d.dataset = make([]uint32, 32*1024/4)
+			generateDataset(d.dataset, d.epoch, cache)
 
 			return
 		}
-		// Full dataset generation is needed, check dataset dir for existing data
+		// If we don't store anything on disk, generate and return
 		csize := cacheSize(d.epoch*epochLength + 1)
 		dsize := datasetSize(d.epoch*epochLength + 1)
 		seed := seedHash(d.epoch*epochLength + 1)
 
-		path := filepath.Join(dir, fmt.Sprintf("full-R%d-%x", algorithmRevision, seed))
-		logger := log.New("seed", hexutil.Bytes(seed))
-
-		if dir != "" {
-			dump, err := os.Open(path)
-			if err == nil {
-				if !discard {
-					logger.Info("Loading ethash DAG from disk")
-					start := time.Now()
-					d.dataset = prepare(dsize, bufio.NewReader(dump))
-					logger.Info("Loaded ethash DAG from disk", "elapsed", common.PrettyDuration(time.Since(start)))
-				}
-				dump.Close()
-				return
-			}
+		if dir == "" {
+			cache := make([]uint32, csize/4)
+			generateCache(cache, d.epoch, seed)
+
+			d.dataset = make([]uint32, dsize/4)
+			generateDataset(d.dataset, d.epoch, cache)
 		}
-		// No previous disk dataset was available, generate on the fly
-		rawCache := generateCache(csize, seed)
-		intCache := prepare(csize, bytes.NewReader(rawCache))
+		// Disk storage is needed, this will get fancy
+		endian := "le"
+		if !isLittleEndian() {
+			endian = "be"
+		}
+		path := filepath.Join(dir, fmt.Sprintf("full-R%d-%x.%s", algorithmRevision, seed, endian))
+		logger := log.New("epoch", d.epoch)
+
+		// Try to load the file from disk and memory map it
+		var err error
+		d.dump, d.mmap, d.dataset, err = memoryMap(path)
+		if err == nil {
+			logger.Debug("Loaded old ethash dataset from disk")
+			return
+		}
+		logger.Debug("Failed to load old ethash dataset", "err", err)
+
+		// No previous dataset available, create a new dataset file to fill
+		cache := make([]uint32, csize/4)
+		generateCache(cache, d.epoch, seed)
 
-		rawDataset := generateDataset(dsize, intCache)
-		if !discard {
-			d.dataset = prepare(dsize, bytes.NewReader(rawDataset))
+		d.dump, d.mmap, d.dataset, err = memoryMapAndGenerate(path, dsize, func(buffer []uint32) { generateDataset(buffer, d.epoch, cache) })
+		if err != nil {
+			logger.Error("Failed to generate mapped ethash dataset", "err", err)
+
+			d.dataset = make([]uint32, dsize/2)
+			generateDataset(d.dataset, d.epoch, cache)
 		}
-		// If a dataset directory is given, attempt to serialize for next time
-		if dir != "" {
-			// Store the ethash dataset to disk
-			start := time.Now()
-			if err := os.MkdirAll(filepath.Dir(path), os.ModePerm); err != nil {
-				logger.Error("Failed to create ethash DAG dir", "err", err)
-			} else if err := ioutil.WriteFile(path, rawDataset, os.ModePerm); err != nil {
-				logger.Error("Failed to write ethash DAG to disk", "err", err)
-			} else {
-				logger.Info("Stored ethash DAG to disk", "elapsed", common.PrettyDuration(time.Since(start)))
-			}
-			// Iterate over all previous instances and delete old ones
-			for ep := int(d.epoch) - limit; ep >= 0; ep-- {
-				seed := seedHash(uint64(ep)*epochLength + 1)
-				path := filepath.Join(dir, fmt.Sprintf("full-R%d-%x", algorithmRevision, seed))
-				os.Remove(path)
-			}
+		// Iterate over all previous instances and delete old ones
+		for ep := int(d.epoch) - limit; ep >= 0; ep-- {
+			seed := seedHash(uint64(ep)*epochLength + 1)
+			path := filepath.Join(dir, fmt.Sprintf("full-R%d-%x.%s", algorithmRevision, seed, endian))
+			os.Remove(path)
 		}
 	})
 }
 
+// release closes any file handlers and memory maps open.
+func (d *dataset) release() {
+	if d.mmap != nil {
+		d.mmap.Unmap()
+		d.mmap = nil
+	}
+	if d.dump != nil {
+		d.dump.Close()
+		d.dump = nil
+	}
+}
+
 // MakeCache generates a new ethash cache and optionally stores it to disk.
 func MakeCache(block uint64, dir string) {
 	c := cache{epoch: block/epochLength + 1}
 	c.generate(dir, math.MaxInt32, false)
+	c.release()
 }
 
 // MakeDataset generates a new ethash dataset and optionally stores it to disk.
 func MakeDataset(block uint64, dir string) {
 	d := dataset{epoch: block/epochLength + 1}
-	d.generate(dir, math.MaxInt32, false, true)
+	d.generate(dir, math.MaxInt32, false)
+	d.release()
 }
 
 // Ethash is a PoW data struture implementing the ethash algorithm.
@@ -318,22 +426,26 @@ func (ethash *Ethash) cache(block uint64) []uint32 {
 				}
 			}
 			delete(ethash.caches, evict.epoch)
+			evict.release()
 
-			log.Debug("Evicted ethash cache", "epoch", evict.epoch, "used", evict.used)
+			log.Trace("Evicted ethash cache", "epoch", evict.epoch, "used", evict.used)
 		}
 		// If we have the new cache pre-generated, use that, otherwise create a new one
 		if ethash.fcache != nil && ethash.fcache.epoch == epoch {
-			log.Debug("Using pre-generated cache", "epoch", epoch)
+			log.Trace("Using pre-generated cache", "epoch", epoch)
 			current, ethash.fcache = ethash.fcache, nil
 		} else {
-			log.Debug("Requiring new ethash cache", "epoch", epoch)
+			log.Trace("Requiring new ethash cache", "epoch", epoch)
 			current = &cache{epoch: epoch}
 		}
 		ethash.caches[epoch] = current
 
 		// If we just used up the future cache, or need a refresh, regenerate
 		if ethash.fcache == nil || ethash.fcache.epoch <= epoch {
-			log.Debug("Requiring new future ethash cache", "epoch", epoch+1)
+			if ethash.fcache != nil {
+				ethash.fcache.release()
+			}
+			log.Trace("Requiring new future ethash cache", "epoch", epoch+1)
 			future = &cache{epoch: epoch + 1}
 			ethash.fcache = future
 		}
@@ -418,23 +530,27 @@ func (ethash *Ethash) dataset(block uint64) []uint32 {
 				}
 			}
 			delete(ethash.datasets, evict.epoch)
+			evict.release()
 
-			log.Debug("Evicted ethash dataset", "epoch", evict.epoch, "used", evict.used)
+			log.Trace("Evicted ethash dataset", "epoch", evict.epoch, "used", evict.used)
 		}
 		// If we have the new cache pre-generated, use that, otherwise create a new one
 		if ethash.fdataset != nil && ethash.fdataset.epoch == epoch {
-			log.Debug("Using pre-generated dataset", "epoch", epoch)
+			log.Trace("Using pre-generated dataset", "epoch", epoch)
 			current = &dataset{epoch: ethash.fdataset.epoch} // Reload from disk
 			ethash.fdataset = nil
 		} else {
-			log.Debug("Requiring new ethash dataset", "epoch", epoch)
+			log.Trace("Requiring new ethash dataset", "epoch", epoch)
 			current = &dataset{epoch: epoch}
 		}
 		ethash.datasets[epoch] = current
 
 		// If we just used up the future dataset, or need a refresh, regenerate
 		if ethash.fdataset == nil || ethash.fdataset.epoch <= epoch {
-			log.Debug("Requiring new future ethash dataset", "epoch", epoch+1)
+			if ethash.fdataset != nil {
+				ethash.fdataset.release()
+			}
+			log.Trace("Requiring new future ethash dataset", "epoch", epoch+1)
 			future = &dataset{epoch: epoch + 1}
 			ethash.fdataset = future
 		}
@@ -443,7 +559,7 @@ func (ethash *Ethash) dataset(block uint64) []uint32 {
 	ethash.lock.Unlock()
 
 	// Wait for generation finish, bump the timestamp and finalize the cache
-	current.generate(ethash.dagdir, ethash.dagsondisk, ethash.tester, false)
+	current.generate(ethash.dagdir, ethash.dagsondisk, ethash.tester)
 
 	current.lock.Lock()
 	current.used = time.Now()
@@ -451,7 +567,7 @@ func (ethash *Ethash) dataset(block uint64) []uint32 {
 
 	// If we exhausted the future dataset, now's a good time to regenerate it
 	if future != nil {
-		go future.generate(ethash.dagdir, ethash.dagsondisk, ethash.tester, true) // Discard results from memorys
+		go future.generate(ethash.dagdir, ethash.dagsondisk, ethash.tester)
 	}
 	return current.dataset
 }
diff --git a/pow/ethash_algo.go b/pow/ethash_algo.go
index d3fac8d5b..ace482b93 100644
--- a/pow/ethash_algo.go
+++ b/pow/ethash_algo.go
@@ -18,15 +18,17 @@ package pow
 
 import (
 	"encoding/binary"
-	"io"
+	"hash"
+	"reflect"
 	"runtime"
 	"sync"
 	"sync/atomic"
 	"time"
+	"unsafe"
 
 	"github.com/ethereum/go-ethereum/common"
-	"github.com/ethereum/go-ethereum/common/hexutil"
 	"github.com/ethereum/go-ethereum/crypto"
+	"github.com/ethereum/go-ethereum/crypto/sha3"
 	"github.com/ethereum/go-ethereum/log"
 )
 
@@ -44,6 +46,22 @@ const (
 	loopAccesses       = 64      // Number of accesses in hashimoto loop
 )
 
+// hasher is a repetitive hasher allowing the same hash data structures to be
+// reused between hash runs instead of requiring new ones to be created.
+type hasher func(dest []byte, data []byte)
+
+// makeHasher creates a repetitive hasher, allowing the same hash data structures
+// to be reused between hash runs instead of requiring new ones to be created.
+//
+// The returned function is not thread safe!
+func makeHasher(h hash.Hash) hasher {
+	return func(dest []byte, data []byte) {
+		h.Write(data)
+		h.Sum(dest[:0])
+		h.Reset()
+	}
+}
+
 // seedHash is the seed to use for generating a verification cache and the mining
 // dataset.
 func seedHash(block uint64) []byte {
@@ -51,9 +69,9 @@ func seedHash(block uint64) []byte {
 	if block < epochLength {
 		return seed
 	}
-	keccak256 := crypto.Keccak256Hasher()
+	keccak256 := makeHasher(sha3.NewKeccak256())
 	for i := 0; i < int(block/epochLength); i++ {
-		seed = keccak256(seed)
+		keccak256(seed, seed)
 	}
 	return seed
 }
@@ -63,17 +81,30 @@ func seedHash(block uint64) []byte {
 // memory, then performing two passes of Sergio Demian Lerner's RandMemoHash
 // algorithm from Strict Memory Hard Hashing Functions (2014). The output is a
 // set of 524288 64-byte values.
-func generateCache(size uint64, seed []byte) []byte {
+//
+// This method places the result into dest in machine byte order.
+func generateCache(dest []uint32, epoch uint64, seed []byte) {
 	// Print some debug logs to allow analysis on low end devices
-	logger := log.New("seed", hexutil.Bytes(seed))
-	logger.Debug("Generating ethash verification cache")
+	logger := log.New("epoch", epoch)
 
 	start := time.Now()
 	defer func() {
-		logger.Info("Generated ethash verification cache", "elapsed", common.PrettyDuration(time.Since(start)))
+		elapsed := time.Since(start)
+
+		logFn := logger.Debug
+		if elapsed > 3*time.Second {
+			logFn = logger.Info
+		}
+		logFn("Generated ethash verification cache", "elapsed", common.PrettyDuration(elapsed))
 	}()
+	// Convert our destination slice to a byte buffer
+	header := *(*reflect.SliceHeader)(unsafe.Pointer(&dest))
+	header.Len *= 4
+	header.Cap *= 4
+	cache := *(*[]byte)(unsafe.Pointer(&header))
 
 	// Calculate the number of thoretical rows (we'll store in one buffer nonetheless)
+	size := uint64(len(cache))
 	rows := int(size) / hashBytes
 
 	// Start a monitoring goroutine to report progress on low end devices
@@ -93,13 +124,12 @@ func generateCache(size uint64, seed []byte) []byte {
 		}
 	}()
 	// Create a hasher to reuse between invocations
-	keccak512 := crypto.Keccak512Hasher()
+	keccak512 := makeHasher(sha3.NewKeccak512())
 
 	// Sequentially produce the initial dataset
-	cache := make([]byte, size)
-	copy(cache, keccak512(seed))
+	keccak512(cache, seed)
 	for offset := uint64(hashBytes); offset < size; offset += hashBytes {
-		copy(cache[offset:], keccak512(cache[offset-hashBytes:offset]))
+		keccak512(cache[offset:], cache[offset-hashBytes:offset])
 		atomic.AddUint32(&progress, 1)
 	}
 	// Use a low-round version of randmemohash
@@ -113,26 +143,31 @@ func generateCache(size uint64, seed []byte) []byte {
 				xorOff = (binary.LittleEndian.Uint32(cache[dstOff:]) % uint32(rows)) * hashBytes
 			)
 			xorBytes(temp, cache[srcOff:srcOff+hashBytes], cache[xorOff:xorOff+hashBytes])
-			copy(cache[dstOff:], keccak512(temp))
+			keccak512(cache[dstOff:], temp)
 
 			atomic.AddUint32(&progress, 1)
 		}
 	}
-	return cache
+	// Swap the byte order on big endian systems and return
+	if !isLittleEndian() {
+		swap(cache)
+	}
+}
+
+// swap changes the byte order of the buffer assuming a uint32 representation.
+func swap(buffer []byte) {
+	for i := 0; i < len(buffer); i += 4 {
+		binary.BigEndian.PutUint32(buffer[i:], binary.LittleEndian.Uint32(buffer[i:]))
+	}
 }
 
 // prepare converts an ethash cache or dataset from a byte stream into the internal
 // int representation. All ethash methods work with ints to avoid constant byte to
 // int conversions as well as to handle both little and big endian systems.
-func prepare(size uint64, r io.Reader) []uint32 {
-	ints := make([]uint32, size/4)
-
-	buffer := make([]byte, 4)
-	for i := 0; i < len(ints); i++ {
-		io.ReadFull(r, buffer)
-		ints[i] = binary.LittleEndian.Uint32(buffer)
+func prepare(dest []uint32, src []byte) {
+	for i := 0; i < len(dest); i++ {
+		dest[i] = binary.LittleEndian.Uint32(src[i*4:])
 	}
-	return ints
 }
 
 // fnv is an algorithm inspired by the FNV hash, which in some cases is used as
@@ -152,7 +187,7 @@ func fnvHash(mix []uint32, data []uint32) {
 
 // generateDatasetItem combines data from 256 pseudorandomly selected cache nodes,
 // and hashes that to compute a single dataset node.
-func generateDatasetItem(cache []uint32, index uint32, keccak512 crypto.Hasher) []byte {
+func generateDatasetItem(cache []uint32, index uint32, keccak512 hasher) []byte {
 	// Calculate the number of thoretical rows (we use one buffer nonetheless)
 	rows := uint32(len(cache) / hashWords)
 
@@ -163,7 +198,7 @@ func generateDatasetItem(cache []uint32, index uint32, keccak512 crypto.Hasher)
 	for i := 1; i < hashWords; i++ {
 		binary.LittleEndian.PutUint32(mix[i*4:], cache[(index%rows)*hashWords+uint32(i)])
 	}
-	mix = keccak512(mix)
+	keccak512(mix, mix)
 
 	// Convert the mix to uint32s to avoid constant bit shifting
 	intMix := make([]uint32, hashWords)
@@ -179,22 +214,39 @@ func generateDatasetItem(cache []uint32, index uint32, keccak512 crypto.Hasher)
 	for i, val := range intMix {
 		binary.LittleEndian.PutUint32(mix[i*4:], val)
 	}
-	return keccak512(mix)
+	keccak512(mix, mix)
+	return mix
 }
 
 // generateDataset generates the entire ethash dataset for mining.
-func generateDataset(size uint64, cache []uint32) []byte {
+//
+// This method places the result into dest in machine byte order.
+func generateDataset(dest []uint32, epoch uint64, cache []uint32) {
 	// Print some debug logs to allow analysis on low end devices
-	logger := log.New("size", size)
-	logger.Debug("Generating ethash dataset")
+	logger := log.New("epoch", epoch)
 
 	defer func(start time.Time) {
-		logger.Debug("Generated ethash dataset", "elapsed", common.PrettyDuration(time.Since(start)))
+		elapsed := time.Since(start)
+
+		logFn := logger.Debug
+		if elapsed > 3*time.Second {
+			logFn = logger.Info
+		}
+		logFn("Generated ethash verification cache", "elapsed", common.PrettyDuration(elapsed))
 	}(time.Now())
 
+	// Figure out whether the bytes need to be swapped for the machine
+	swapped := !isLittleEndian()
+
+	// Convert our destination slice to a byte buffer
+	header := *(*reflect.SliceHeader)(unsafe.Pointer(&dest))
+	header.Len *= 4
+	header.Cap *= 4
+	dataset := *(*[]byte)(unsafe.Pointer(&header))
+
 	// Generate the dataset on many goroutines since it takes a while
-	dataset := make([]byte, size)
 	threads := runtime.NumCPU()
+	size := uint64(len(dataset))
 
 	var pend sync.WaitGroup
 	pend.Add(threads)
@@ -205,7 +257,7 @@ func generateDataset(size uint64, cache []uint32) []byte {
 			defer pend.Done()
 
 			// Create a hasher to reuse between invocations
-			keccak512 := crypto.Keccak512Hasher()
+			keccak512 := makeHasher(sha3.NewKeccak512())
 
 			// Calculate the data segment this thread should generate
 			batch := uint32(size / hashBytes / uint64(threads))
@@ -217,7 +269,11 @@ func generateDataset(size uint64, cache []uint32) []byte {
 			// Calculate the dataset segment
 			percent := uint32(size / hashBytes / 100)
 			for index := start; index < limit; index++ {
-				copy(dataset[index*hashBytes:], generateDatasetItem(cache, index, keccak512))
+				item := generateDatasetItem(cache, index, keccak512)
+				if swapped {
+					swap(item)
+				}
+				copy(dataset[index*hashBytes:], item)
 
 				if status := atomic.AddUint32(&progress, 1); status%percent == 0 {
 					logger.Info("Generating DAG in progress", "percentage", uint64(status*100)/(size/hashBytes))
@@ -227,8 +283,6 @@ func generateDataset(size uint64, cache []uint32) []byte {
 	}
 	// Wait for all the generators to finish and return
 	pend.Wait()
-
-	return dataset
 }
 
 // hashimoto aggregates data from the full dataset in order to produce our final
@@ -277,7 +331,7 @@ func hashimoto(hash []byte, nonce uint64, size uint64, lookup func(index uint32)
 // in-memory cache) in order to produce our final value for a particular header
 // hash and nonce.
 func hashimotoLight(size uint64, cache []uint32, hash []byte, nonce uint64) ([]byte, []byte) {
-	keccak512 := crypto.Keccak512Hasher()
+	keccak512 := makeHasher(sha3.NewKeccak512())
 
 	lookup := func(index uint32) []uint32 {
 		rawData := generateDatasetItem(cache, index, keccak512)
diff --git a/pow/ethash_algo_test.go b/pow/ethash_algo_test.go
index 32e115db9..c881874ff 100644
--- a/pow/ethash_algo_test.go
+++ b/pow/ethash_algo_test.go
@@ -18,21 +18,28 @@ package pow
 
 import (
 	"bytes"
+	"io/ioutil"
+	"math/big"
+	"os"
+	"reflect"
+	"sync"
 	"testing"
 
+	"github.com/ethereum/go-ethereum/common"
 	"github.com/ethereum/go-ethereum/common/hexutil"
+	"github.com/ethereum/go-ethereum/core/types"
 )
 
 // Tests that verification caches can be correctly generated.
 func TestCacheGeneration(t *testing.T) {
 	tests := []struct {
 		size  uint64
-		seed  []byte
+		epoch uint64
 		cache []byte
 	}{
 		{
-			size: 1024,
-			seed: make([]byte, 32),
+			size:  1024,
+			epoch: 0,
 			cache: hexutil.MustDecode("0x" +
 				"7ce2991c951f7bf4c4c1bb119887ee07871eb5339d7b97b8588e85c742de90e5bafd5bbe6ce93a134fb6be9ad3e30db99d9528a2ea7846833f52e9ca119b6b54" +
 				"8979480c46e19972bd0738779c932c1b43e665a2fd3122fc3ddb2691f353ceb0ed3e38b8f51fd55b6940290743563c9f8fa8822e611924657501a12aafab8a8d" +
@@ -52,8 +59,8 @@ func TestCacheGeneration(t *testing.T) {
 				"845f64fd8324bb85312979dead74f764c9677aab89801ad4f927f1c00f12e28f22422bb44200d1969d9ab377dd6b099dc6dbc3222e9321b2c1e84f8e2f07731c"),
 		},
 		{
-			size: 1024,
-			seed: hexutil.MustDecode("0x290decd9548b62a8d60345a988386fc84ba6bc95484008f6362f93160ef3e563"),
+			size:  1024,
+			epoch: 1,
 			cache: hexutil.MustDecode("0x" +
 				"1f56855d59cc5a085720899b4377a0198f1abe948d85fe5820dc0e346b7c0931b9cde8e541d751de3b2b3275d0aabfae316209d5879297d8bd99f8a033c9d4df" +
 				"35add1029f4e6404a022d504fb8023e42989aba985a65933b0109c7218854356f9284983c9e7de97de591828ae348b63d1fc78d8db58157344d4e06530ffd422" +
@@ -74,22 +81,28 @@ func TestCacheGeneration(t *testing.T) {
 		},
 	}
 	for i, tt := range tests {
-		if cache := generateCache(tt.size, tt.seed); !bytes.Equal(cache, tt.cache) {
-			t.Errorf("cache %d: content mismatch: have %x, want %x", i, cache, tt.cache)
+		cache := make([]uint32, tt.size/4)
+		generateCache(cache, tt.epoch, seedHash(tt.epoch*epochLength+1))
+
+		want := make([]uint32, tt.size/4)
+		prepare(want, tt.cache)
+
+		if !reflect.DeepEqual(cache, want) {
+			t.Errorf("cache %d: content mismatch: have %x, want %x", i, cache, want)
 		}
 	}
 }
 
 func TestDatasetGeneration(t *testing.T) {
 	tests := []struct {
+		epoch       uint64
 		cacheSize   uint64
-		cacheSeed   []byte
 		datasetSize uint64
 		dataset     []byte
 	}{
 		{
+			epoch:       0,
 			cacheSize:   1024,
-			cacheSeed:   make([]byte, 32),
 			datasetSize: 32 * 1024,
 			dataset: hexutil.MustDecode("0x" +
 				"4bc09fbd530a041dd2ec296110a29e8f130f179c59d223f51ecce3126e8b0c74326abc2f32ccd9d7f976bd0944e3ccf8479db39343cbbffa467046ca97e2da63" +
@@ -608,11 +621,17 @@ func TestDatasetGeneration(t *testing.T) {
 		},
 	}
 	for i, tt := range tests {
-		rawCache := generateCache(tt.cacheSize, tt.cacheSeed)
-		cache := prepare(uint64(len(rawCache)), bytes.NewReader(rawCache))
+		cache := make([]uint32, tt.cacheSize/4)
+		generateCache(cache, tt.epoch, seedHash(tt.epoch*epochLength+1))
+
+		dataset := make([]uint32, tt.datasetSize/4)
+		generateDataset(dataset, tt.epoch, cache)
 
-		if dataset := generateDataset(tt.datasetSize, cache); !bytes.Equal(dataset, tt.dataset) {
-			t.Errorf("dataset %d: content mismatch: have %x, want %x", i, dataset, tt.dataset)
+		want := make([]uint32, tt.datasetSize/4)
+		prepare(want, tt.dataset)
+
+		if !reflect.DeepEqual(dataset, want) {
+			t.Errorf("dataset %d: content mismatch: have %x, want %x", i, dataset, want)
 		}
 	}
 }
@@ -621,12 +640,12 @@ func TestDatasetGeneration(t *testing.T) {
 // datasets.
 func TestHashimoto(t *testing.T) {
 	// Create the verification cache and mining dataset
-	var (
-		rawCache   = generateCache(1024, make([]byte, 32))
-		cache      = prepare(uint64(len(rawCache)), bytes.NewReader(rawCache))
-		rawDataset = generateDataset(32*1024, cache)
-		dataset    = prepare(uint64(len(rawDataset)), bytes.NewReader(rawDataset))
-	)
+	cache := make([]uint32, 1024/4)
+	generateCache(cache, 0, make([]byte, 32))
+
+	dataset := make([]uint32, 32*1024/4)
+	generateDataset(dataset, 0, cache)
+
 	// Create a block to verify
 	hash := hexutil.MustDecode("0xc9149cc0386e689d789a1c2f3d5d169a61a6218ed30e74414dc736e442ef3d1f")
 	nonce := uint64(0)
@@ -650,31 +669,77 @@ func TestHashimoto(t *testing.T) {
 	}
 }
 
+// Tests that caches generated on disk may be done concurrently.
+func TestConcurrentDiskCacheGeneration(t *testing.T) {
+	// Create a temp folder to generate the caches into
+	cachedir, err := ioutil.TempDir("", "")
+	if err != nil {
+		t.Fatalf("Failed to create temporary cache dir: %v", err)
+	}
+	defer os.RemoveAll(cachedir)
+
+	// Define a heavy enough block, one from mainnet should do
+	block := types.NewBlockWithHeader(&types.Header{
+		Number:      big.NewInt(3311058),
+		ParentHash:  common.HexToHash("0xd783efa4d392943503f28438ad5830b2d5964696ffc285f338585e9fe0a37a05"),
+		UncleHash:   common.HexToHash("0x1dcc4de8dec75d7aab85b567b6ccd41ad312451b948a7413f0a142fd40d49347"),
+		Coinbase:    common.HexToAddress("0xc0ea08a2d404d3172d2add29a45be56da40e2949"),
+		Root:        common.HexToHash("0x77d14e10470b5850332524f8cd6f69ad21f070ce92dca33ab2858300242ef2f1"),
+		TxHash:      common.HexToHash("0x56e81f171bcc55a6ff8345e692c0f86e5b48e01b996cadc001622fb5e363b421"),
+		ReceiptHash: common.HexToHash("0x56e81f171bcc55a6ff8345e692c0f86e5b48e01b996cadc001622fb5e363b421"),
+		Difficulty:  big.NewInt(167925187834220),
+		GasLimit:    big.NewInt(4015682),
+		GasUsed:     big.NewInt(0),
+		Time:        big.NewInt(1488928920),
+		Extra:       []byte("www.bw.com"),
+		MixDigest:   common.HexToHash("0x3e140b0784516af5e5ec6730f2fb20cca22f32be399b9e4ad77d32541f798cd0"),
+		Nonce:       types.EncodeNonce(0xf400cd0006070c49),
+	})
+	// Simulate multiple processes sharing the same datadir
+	var pend sync.WaitGroup
+
+	for i := 0; i < 3; i++ {
+		pend.Add(1)
+
+		go func(idx int) {
+			defer pend.Done()
+
+			ethash := NewFullEthash(cachedir, 0, 1, "", 0, 0)
+			if err := ethash.Verify(block); err != nil {
+				t.Errorf("proc %d: block verification failed: %v", idx, err)
+			}
+		}(i)
+	}
+	pend.Wait()
+}
+
 // Benchmarks the cache generation performance.
 func BenchmarkCacheGeneration(b *testing.B) {
 	for i := 0; i < b.N; i++ {
-		generateCache(cacheSize(1), make([]byte, 32))
+		cache := make([]uint32, cacheSize(1)/4)
+		generateCache(cache, 0, make([]byte, 32))
 	}
 }
 
 // Benchmarks the dataset (small) generation performance.
 func BenchmarkSmallDatasetGeneration(b *testing.B) {
-	rawCache := generateCache(65536, make([]byte, 32))
-	cache := prepare(uint64(len(rawCache)), bytes.NewReader(rawCache))
+	cache := make([]uint32, 65536/4)
+	generateCache(cache, 0, make([]byte, 32))
 
 	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
-		generateDataset(32*65536, cache)
+		dataset := make([]uint32, 32*65536/4)
+		generateDataset(dataset, 0, cache)
 	}
 }
 
 // Benchmarks the light verification performance.
 func BenchmarkHashimotoLight(b *testing.B) {
-	var (
-		rawCache = generateCache(cacheSize(1), make([]byte, 32))
-		cache    = prepare(uint64(len(rawCache)), bytes.NewReader(rawCache))
-		hash     = hexutil.MustDecode("0xc9149cc0386e689d789a1c2f3d5d169a61a6218ed30e74414dc736e442ef3d1f")
-	)
+	cache := make([]uint32, cacheSize(1)/4)
+	generateCache(cache, 0, make([]byte, 32))
+
+	hash := hexutil.MustDecode("0xc9149cc0386e689d789a1c2f3d5d169a61a6218ed30e74414dc736e442ef3d1f")
+
 	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
 		hashimotoLight(datasetSize(1), cache, hash, 0)
@@ -683,13 +748,14 @@ func BenchmarkHashimotoLight(b *testing.B) {
 
 // Benchmarks the full (small) verification performance.
 func BenchmarkHashimotoFullSmall(b *testing.B) {
-	var (
-		rawCache   = generateCache(65536, make([]byte, 32))
-		cache      = prepare(uint64(len(rawCache)), bytes.NewReader(rawCache))
-		rawDataset = generateDataset(32*65536, cache)
-		dataset    = prepare(uint64(len(rawDataset)), bytes.NewReader(rawDataset))
-		hash       = hexutil.MustDecode("0xc9149cc0386e689d789a1c2f3d5d169a61a6218ed30e74414dc736e442ef3d1f")
-	)
+	cache := make([]uint32, 65536/4)
+	generateCache(cache, 0, make([]byte, 32))
+
+	dataset := make([]uint32, 32*65536/4)
+	generateDataset(dataset, 0, cache)
+
+	hash := hexutil.MustDecode("0xc9149cc0386e689d789a1c2f3d5d169a61a6218ed30e74414dc736e442ef3d1f")
+
 	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
 		hashimotoFull(32*65536, dataset, hash, 0)