diff --git a/perffile/bufdecoder.go b/perffile/bufdecoder.go
index fcb5966..fa4fbb0 100644
--- a/perffile/bufdecoder.go
+++ b/perffile/bufdecoder.go
@@ -45,6 +45,10 @@ func (b *bufDecoder) i32() int32 {
 }
 
 func (b *bufDecoder) u64() uint64 {
+	// see https://github.com/aclements/go-perf/issues/5
+	if len(b.buf) < 8 {
+		return 0
+	}
 	x := b.order.Uint64(b.buf)
 	b.buf = b.buf[8:]
 	return x
diff --git a/perffile/reader.go b/perffile/reader.go
index 866f259..19be18b 100644
--- a/perffile/reader.go
+++ b/perffile/reader.go
@@ -356,10 +356,10 @@ func (f *File) Records(order RecordsOrder) *Records {
 			return &Records{err: rs.Err()}
 		}
 		sort.Stable(&timeSorter{pos, ts})
-		return &Records{f: f, sr: newBufferedSectionReader(f.hdr.Data.sectionReader(f.r)), order: pos}
+		return &Records{f: f, sr: newBufferedSectionReader(f.hdr.Data.sectionReader(f.r)), order: pos, hdrBuf: make([]byte, 8)}
 	}
 
-	return &Records{f: f, sr: newBufferedSectionReader(f.hdr.Data.sectionReader(f.r))}
+	return &Records{f: f, sr: newBufferedSectionReader(f.hdr.Data.sectionReader(f.r)), hdrBuf: make([]byte, 8)}
 }
 
 type timeSorter struct {
diff --git a/perffile/records.go b/perffile/records.go
index bafba8c..e8c0714 100644
--- a/perffile/records.go
+++ b/perffile/records.go
@@ -50,6 +50,10 @@ type Records struct {
 	recordAux           RecordAux
 	recordSwitch        RecordSwitch
 	recordSwitchCPUWide RecordSwitchCPUWide
+
+	// reuse of recordHeader to reduce memory alloc objects
+	hdr    recordHeader
+	hdrBuf []byte
 }
 
 // Err returns the first error encountered by Records.
@@ -57,6 +61,16 @@ func (r *Records) Err() error {
 	return r.err
 }
 
+func (r *Records) readRecordHeader() error {
+	if _, err := io.ReadFull(r.sr, r.hdrBuf); err != nil {
+		return err
+	}
+	r.hdr.Type = RecordType(binary.LittleEndian.Uint32(r.hdrBuf))
+	r.hdr.Misc = recordMisc(binary.LittleEndian.Uint16(r.hdrBuf[4:]))
+	r.hdr.Size = binary.LittleEndian.Uint16(r.hdrBuf[6:])
+	return nil
+}
+
 // Next fetches the next record into r.Record.  It returns true if
 // successful, and false if it reaches the end of the record stream or
 // encounters an error.
@@ -87,8 +101,7 @@ func (r *Records) Next() bool {
 	common.Offset = offset + int64(r.f.hdr.Data.Offset)
 
 	// Read record header
-	var hdr recordHeader
-	if err := binary.Read(r.sr, binary.LittleEndian, &hdr); err != nil {
+	if err := r.readRecordHeader(); err != nil {
 		if err != io.EOF {
 			r.err = err
 		}
@@ -96,7 +109,7 @@ func (r *Records) Next() bool {
 	}
 
 	// Read record data
-	rlen := int(hdr.Size - 8)
+	rlen := int(r.hdr.Size - 8)
 	if rlen > len(r.buf) {
 		r.buf = make([]byte, rlen)
 	}
@@ -107,90 +120,90 @@ func (r *Records) Next() bool {
 	}
 
 	// Parse common sample_id fields
-	if r.f.sampleIDAll && hdr.Type != RecordTypeSample && hdr.Type < recordTypeUserStart {
+	if r.f.sampleIDAll && r.hdr.Type != RecordTypeSample && r.hdr.Type < recordTypeUserStart {
 		// mmap records in the prologue don't have eventAttrs
 		// in recent perf versions, but that's okay.
 		//
 		// TODO: When is perf okay with missing eventAttrs?
-		r.parseCommon(bd, &common, hdr.Type == RecordTypeMmap)
+		r.parseCommon(bd, &common, r.hdr.Type == RecordTypeMmap)
 	}
 
 	// Parse record
 	// TODO: Don't array out-of-bounds on short records
-	switch hdr.Type {
+	switch r.hdr.Type {
 	default:
 		// As far as I can tell, RecordTypeRead can never
 		// appear in a perf.data file.
-		r.Record = &RecordUnknown{hdr, common, bd.buf}
+		r.Record = &RecordUnknown{r.hdr, common, bd.buf}
 
 	case RecordTypeMmap:
-		r.Record = r.parseMmap(bd, &hdr, &common, false)
+		r.Record = r.parseMmap(bd, &r.hdr, &common, false)
 
 	case RecordTypeLost:
-		r.Record = r.parseLost(bd, &hdr, &common)
+		r.Record = r.parseLost(bd, &r.hdr, &common)
 
 	case RecordTypeComm:
-		r.Record = r.parseComm(bd, &hdr, &common)
+		r.Record = r.parseComm(bd, &r.hdr, &common)
 
 	case RecordTypeExit:
-		r.Record = r.parseExit(bd, &hdr, &common)
+		r.Record = r.parseExit(bd, &r.hdr, &common)
 
 	case RecordTypeThrottle:
-		r.Record = r.parseThrottle(bd, &hdr, &common, true)
+		r.Record = r.parseThrottle(bd, &r.hdr, &common, true)
 
 	case RecordTypeUnthrottle:
-		r.Record = r.parseThrottle(bd, &hdr, &common, false)
+		r.Record = r.parseThrottle(bd, &r.hdr, &common, false)
 
 	case RecordTypeFork:
-		r.Record = r.parseFork(bd, &hdr, &common)
+		r.Record = r.parseFork(bd, &r.hdr, &common)
 
 	case RecordTypeSample:
-		r.Record = r.parseSample(bd, &hdr, &common)
+		r.Record = r.parseSample(bd, &r.hdr, &common)
 
 	case recordTypeMmap2:
-		r.Record = r.parseMmap(bd, &hdr, &common, true)
+		r.Record = r.parseMmap(bd, &r.hdr, &common, true)
 
 	case RecordTypeAux:
-		r.Record = r.parseAux(bd, &hdr, &common)
+		r.Record = r.parseAux(bd, &r.hdr, &common)
 
 	case RecordTypeItraceStart:
-		r.Record = r.parseItraceStart(bd, &hdr, &common)
+		r.Record = r.parseItraceStart(bd, &r.hdr, &common)
 
 	case RecordTypeLostSamples:
-		r.Record = r.parseLostSamples(bd, &hdr, &common)
+		r.Record = r.parseLostSamples(bd, &r.hdr, &common)
 
 	case RecordTypeSwitch:
-		r.Record = r.parseSwitch(bd, &hdr, &common)
+		r.Record = r.parseSwitch(bd, &r.hdr, &common)
 
 	case RecordTypeSwitchCPUWide:
-		r.Record = r.parseSwitchCPUWide(bd, &hdr, &common)
+		r.Record = r.parseSwitchCPUWide(bd, &r.hdr, &common)
 
 	case RecordTypeNamespaces:
-		r.Record = r.parseNamespaces(bd, &hdr, &common)
+		r.Record = r.parseNamespaces(bd, &r.hdr, &common)
 
 	case RecordTypeKsymbol:
-		r.Record = r.parseKsymbol(bd, &hdr, &common)
+		r.Record = r.parseKsymbol(bd, &r.hdr, &common)
 
 	case RecordTypeBPFEvent:
-		r.Record = r.parseBPFEvent(bd, &hdr, &common)
+		r.Record = r.parseBPFEvent(bd, &r.hdr, &common)
 
 	case RecordTypeCGroup:
-		r.Record = r.parseCGroup(bd, &hdr, &common)
+		r.Record = r.parseCGroup(bd, &r.hdr, &common)
 
 	case RecordTypeTextPoke:
-		r.Record = r.parseTextPoke(bd, &hdr, &common)
+		r.Record = r.parseTextPoke(bd, &r.hdr, &common)
 
 	case RecordTypeAuxOutputHardwareID:
-		r.Record = r.parseAuxOutputHardwareID(bd, &hdr, &common)
+		r.Record = r.parseAuxOutputHardwareID(bd, &r.hdr, &common)
 
 	case RecordTypeAuxtraceInfo:
-		r.Record = r.parseAuxtraceInfo(bd, &hdr, &common)
+		r.Record = r.parseAuxtraceInfo(bd, &r.hdr, &common)
 
 	case RecordTypeAuxtrace:
 		// Note: This appears to be the only record type that
 		// has additional payload data following it that isn't
 		// included in the header size.
-		r.Record = r.parseAuxtrace(bd, &hdr, &common)
+		r.Record = r.parseAuxtrace(bd, &r.hdr, &common)
 	}
 	if r.err != nil {
 		return false