1. In-memory Caching in HDFS: Lower
Latency, Same Great Taste
Andrew Wang and Colin McCabe
Cloudera

2. 2
In-memory Caching in HDFS
Lower latency, same great taste
Andrew Wang | awang@cloudera.com
Colin McCabe | cmccabe@cloudera.com

3. Alice
Hadoop
cluster
Query
Result set

4. Alice
Fresh data

5. Fresh data

6. Alice
Rollup

7. Problems
• Data hotspots
• Everyone wants to query some fresh data
• Shared disks are unable to handle high load
• Mixed workloads
• Data analyst making small point queries
• Rollup job scanning all the data
• Point query latency suffers because of I/O contention
• Same theme: disk I/O contention!
7

8. How do we solve I/O issues?
• Cache important datasets in memory!
• Much higher throughput than disk
• Fast random/concurrent access
• Interesting working sets often fit in cluster memory
• Traces from Facebook’s Hive cluster
• Increasingly affordable to buy a lot of memory
• Moore’s law
• 1TB RAM server is 40k on HP’s website
8

9. Alice
Page cache

10. Alice
Repeated
query
?

11. Alice
Rollup

12. Alice
Extra copies

13. Alice
Checksum
verification
Extra copies

14. Design Considerations
1. Placing tasks for memory locality
• Expose cache locations to application schedulers
2. Contention for page cache from other users
• Explicitly pin hot datasets in memory
3. Extra copies when reading cached data
• Zero-copy API to read cached data
14

15. Outline
• Implementation
• NameNode and DataNode modifications
• Zero-copy read API
• Evaluation
• Microbenchmarks
• MapReduce
• Impala
• Future work
15

16. Outline
• Implementation
• NameNode and DataNode modifications
• Zero-copy read API
• Evaluation
• Microbenchmarks
• MapReduce
• Impala
• Future work
16

17. Cache Directives
• A cache directive describes a file or directory that
should be cached
• Path
• Cache replication factor: 1-N
• Stored permanently on the NameNode
• Also have cache pools for access control and quotas,
but we won’t be covering that here
17

18. Architecture
18
DataNode
DataNode DataNode
NameNode
DFSClient
Cache /foo
Cache commandsCache Heartbeats
DFSClient
getBlockLocations

19. mlock
• The DataNode pins each cached block into the page
cache using mlock, and checksums it.
• Because we’re using the page cache, the blocks don’t
take up any space on the Java heap.
19
DataNode
Page Cache
DFSClient read
mlock

20. Zero-copy read API
• Clients can use the zero-copy read API to map the
cached replica into their own address space
• The zero-copy API avoids the overhead of the
read() and pread() system calls
• However, we don’t verify checksums when using the
zero-copy API
• The zero-copy API can be only used on cached data, or
when the application computes its own checksums.
20

21. Zero-copy read API
New FSDataInputStream methods:
ByteBuffer read(ByteBufferPool pool,
int maxLength, EnumSet<ReadOption> opts);
void releaseBuffer(ByteBuffer buffer);
21

22. Skipping Checksums
• We would like to skip checksum verification when
reading cached data
• DataNode already checksums when caching the block
• Enables more efficient SCR, ZCR
• Requirements
• Client needs to know that the replica is cached
• DataNode needs to notify the client if the replica is
uncached
22

23. Skipping Checksums
• The DataNode and DFSClient use shared memory
segments to communicate which blocks are cached.
23
DataNode
Page Cache
DFSClient read
mlock
Shared
Memory
Segment

24. Skipping Checksums
24
Block 123
DataNode DFSClient
Can Skip Csums
In Use
Zero-Copy
MappedByteBuffer

25. Skipping Checksums
25
Block 123
DataNode DFSClient
Can Skip Csums
In Use
Zero-Copy
MappedByteBuffer

26. Architecture Summary
• The Cache Directive API provides per-file control over
what is cached
• The NameNode tracks cached blocks and coordinates
DataNode cache work
• The DataNodes use mlock to lock page cache blocks
into memory
• The DFSClient can determine whether it is safe to
skip checksums via the shared memory segment
• Caching makes it possible to use the efficient Zero-
Copy API on cached data
26

27. Outline
• Implementation
• NameNode and DataNode modifications
• Zero-copy read API
• Evaluation
• Single-Node Microbenchmarks
• MapReduce
• Impala
• Future work
27

28. Test Node
• 48GB of RAM
• Configured 38GB of HDFS cache
• 11x SATA hard disks
• 2x4 core 2.13 GHz Westmere Xeon processors
• 10 Gbit/s full-bisection bandwidth network
28

29. Single-Node Microbenchmarks
• How much faster are cached and zero-copy reads?
• Introducing vecsum (vector sum)
• Computes sums of a file of doubles
• Highly optimized: uses SSE intrinsics
• libhdfs program
• Can toggle between various read methods
• Terminology
• SCR: short-circuit reads
• ZCR: zero-copy reads
29

30. Throughput Reading 1G File 20x
30
0.8 0.9
1.9
2.4
5.9
0
1
2
3
4
5
6
7
TCP TCP no
csums
SCR SCR no
csums
ZCR
GB/s

31. ZCR 1GB vs 20GB
31
5.9
2.7
0
1
2
3
4
5
6
7
1GB 20GB
GB/s

32. Throughput
• Skipping checksums matters more when going faster
• ZCR gets close to bus bandwidth
• ~6GB/s
• Need to reuse client-side mmaps for maximum perf
• page_fault function is 1.16% of cycles in 1G
• 17.55% in 20G
32

33. Client CPU cycles
33
57.6
51.8
27.1
23.4
12.7
0
10
20
30
40
50
60
70
TCP TCP no
csums
SCR SCR no
csums
ZCR
CPUcycles(billions)

34. Why is ZCR more CPU-efficient?
34

35. Why is ZCR more CPU-efficient?
35

36. Remote Cached vs. Local Uncached
• Zero-copy is only possible for local cached data
• Is it better to read from remote cache, or local disk?
36

37. Remote Cached vs. Local Uncached
37
841
1092
125 137
0
200
400
600
800
1000
1200
TCP iperf SCR dd
MB/s

38. Microbenchmark Conclusions
• Short-circuit reads need less CPU than TCP reads
• ZCR is even more efficient, because it avoids a copy
• ZCR goes much faster when re-reading the same
data, because it can avoid mmap page faults
• Network and disk may be bottleneck for remote or
uncached reads
38

39. Outline
• Implementation
• NameNode and DataNode modifications
• Zero-copy read API
• Evaluation
• Microbenchmarks
• MapReduce
• Impala
• Future work
39

40. MapReduce
• Started with example MR jobs
• Wordcount
• Grep
• 5 node cluster: 4 DNs, 1 NN
• Same hardware configuration as single node tests
• 38GB HDFS cache per DN
• 11 disks per DN
• 17GB of Wikipedia text
• Small enough to fit into cache at 3x replication
• Ran each job 10 times, took the average
40

41. wordcount and grep
41
275
52
280
55
0
50
100
150
200
250
300
350
400
wordcount wordcount
cached
grep grep cached

42. wordcount and grep
42
275
52
280
55
0
50
100
150
200
250
300
350
400
wordcount wordcount
cached
grep grep cached
Almost no
speedup!

43. wordcount and grep
43
275
52
280
55
0
50
100
150
200
250
300
350
400
wordcount wordcount
cached
grep grep cached
~60MB/s
~330MB/s
Not I/O bound

44. wordcount and grep
• End-to-end latency barely changes
• These MR jobs are simply not I/O bound!
• Best map phase throughput was about 330MB/s
• 44 disks can theoretically do 4400MB/s
• Further reasoning
• Long JVM startup and initialization time
• Many copies in TextInputFormat, doesn’t use zero-copy
• Caching input data doesn’t help reduce step
44

45. Introducing bytecount
• Trivial version of wordcount
• Counts # of occurrences of byte values
• Heavily CPU optimized
• Each mapper processes an entire block via ZCR
• No additional copies
• No record slop across block boundaries
• Fast inner loop
• Very unrealistic job, but serves as a best case
• Also tried 2GB block size to amortize startup costs
45

46. bytecount
46
52
39 35
55
45
58
0
10
20
30
40
50
60
70

47. bytecount
47
52
39 35
55
45
58
0
10
20
30
40
50
60
70
1.3x faster

48. bytecount
48
52
39 35
55
45
58
0
10
20
30
40
50
60
70
Still only
~500MB/s

49. MapReduce Conclusions
49
• Many MR jobs will see marginal improvement
• Startup costs
• CPU inefficiencies
• Shuffle and reduce steps
• Even bytecount sees only modest gains
• 1.3x faster than disk
• 500MB/s with caching and ZCR
• Nowhere close to GB/s possible with memory
• Needs more work to take full advantage of caching!

50. Outline
• Implementation
• NameNode and DataNode modifications
• Zero-copy read API
• Evaluation
• Microbenchmarks
• MapReduce
• Impala
• Future work
50

51. Impala Benchmarks
• Open-source OLAP database developed by Cloudera
• Tested with Impala 1.3 (CDH 5)
• Same 4 DN cluster as MR section
• 38GB of 48GB per DN configured as HDFS cache
• 152GB aggregate HDFS cache
• 11 disks per DN
51

52. Impala Benchmarks
• 1TB TPC-DS store_sales table, text format
• count(*) on different numbers of partitions
• Has to scan all the data, no skipping
• Queries
• 51GB small query (34% cache capacity)
• 148GB big query (98% cache capacity)
• Small query with concurrent workload
• Tested “cold” and “hot”
• echo 3 > /proc/sys/vm/drop_caches
• Lets us compare HDFS caching against page cache
52

53. Small Query
53
19.8
5.8
4.0 3.0
0
5
10
15
20
25
Uncached
cold
Cached cold Uncached
hot
Cached hot
Averageresponsetime(s)

54. Small Query
54
19.8
5.8
4.0 3.0
0
5
10
15
20
25
Uncached
cold
Cached cold Uncached
hot
Cached hot
Averageresponsetime(s)
2550 MB/s 17 GB/s
I/O bound!

55. Small Query
55
0
5
10
15
20
25
Uncached
cold
Cached cold Uncached
hot
Cached hot
Averageresponsetime(s)
3.4x faster,
disk vs. memory

56. Small Query
56
0
5
10
15
20
25
Uncached
cold
Cached cold Uncached
hot
Cached hot
Averageresponsetime(s)
1.3x after warmup, still
wins on CPU efficiency

57. Big Query
57
48.2
11.5
40.9
9.4
0
10
20
30
40
50
60
Uncached
cold
Cached cold Uncached
hot
Cached hot
Averageresponsetime(s)

58. Big Query
58
0
10
20
30
40
50
60
Uncached
cold
Cached cold Uncached
hot
Cached hot
Averageresponsetime(s)
4.2x
faster, disk
vs mem

59. Big Query
59
0
10
20
30
40
50
60
Uncached
cold
Cached cold Uncached
hot
Cached hot
Averageresponsetime(s)
4.3x
faster, does
n’t fit in
page cache
Cannot schedule for
page cache locality

60. Small Query with Concurrent Workload
60
0
10
20
30
40
50
60
Uncached Cached Cached (not
concurrent)
Averageresponsetime(s)

61. Small Query with Concurrent Workload
61
0
10
20
30
40
50
60
Uncached Cached Cached (not
concurrent)
Averageresponsetime(s)
7x faster when small query
working set is cached

62. Small Query with Concurrent Workload
62
0
10
20
30
40
50
60
Uncached Cached Cached (not
concurrent)
Averageresponsetime(s)
2x slower than
isolated, CPU
contention

63. Impala Conclusions
• HDFS cache is faster than disk or page cache
• ZCR is more efficient than SCR from page cache
• Better when working set is approx. cluster memory
• Can schedule tasks for cache locality
• Significantly better for concurrent workloads
• 7x faster when contending with a single background query
• Impala performance will only improve
• Many CPU improvements on the roadmap
63

64. Outline
• Implementation
• NameNode and DataNode modifications
• Zero-copy read API
• Evaluation
• Microbenchmarks
• MapReduce
• Impala
• Future work
64

65. Future Work
• Automatic cache replacement
• LRU, LFU, ?
• Sub-block caching
• Potentially important for automatic cache replacement
• Columns in Parquet
• Compression, encryption, serialization
• Lose many benefits of zero-copy API
• Write-side caching
• Enables Spark-like RDDs for all HDFS applications
65

66. Conclusion
• I/O contention is a problem for concurrent workloads
• HDFS can now explicitly pin working sets into RAM
• Applications can place their tasks for cache locality
• Use zero-copy API to efficiently read cached data
• Substantial performance improvements
• 6GB/s for single thread microbenchmark
• 7x faster for concurrent Impala workload
66

70. bytecount
70
52
39 35
55
45
58
0
10
20
30
40
50
60
70
Less disk parallelism