Presentation slides from EuroSys 2007.

1. Thread Clustering:
Sharing-Aware Thread Scheduling
on SMP-CMP-SMT Multiprocessors
David Tam, Reza Azimi, Michael Stumm
University of Toronto
{tamda, azimi, stumm}@eecg.toronto.edu
Thread Clustering

2. Multiprocessors Today
Example: IBM Power 5 system
1
Thread Clustering

3. Multiprocessors Today
Example: IBM Power 5 system
SMP
CMP
SMT
SHARED
CACHE
1
Thread Clustering

4. Multiprocessors Today
Example: IBM Power 5 system
Disparity in L2 latencies
120
14
1
Thread Clustering

5. Operating Systems Today
CPU Schedulers:
Ignore disparity in L2 latencies
●
Ignore data sharing among threads
●
Distribute threads poorly
●
Cross-chip traffic
●
Remote L2 cache accesses
●
Causes performance problem
●
2
Thread Clustering

6. Our Goal: Sharing-Aware Scheduling
Detect sharing patterns
●
Cluster threads
●
Benefits:
Decrease cross-chip traffic
●
Increase on-chip cache locality
●
Exploit shared L2 caches
●
3
Thread Clustering

7. Our Online Technique
STEPS:
1) Monitor remote cache access rate
2) Detect thread sharing patterns
REPEAT
3) Determine thread clusters
4) Migrate thread clusters
4
Thread Clustering

8. Sharing Detection
To observe remote cache accesses:
●
Exploit HPCs (hardware performance counters)
●
Sample remote cache miss addresses
●
Local cache misses satisfied by remote cache
●
IBM Power 5 continuous data sampling
●
1
X
5
Thread Clustering

9. Sharing Detection
To observe remote cache accesses:
●
Exploit HPCs (hardware performance counters)
●
Sample remote cache miss addresses
●
Local cache misses satisfied by remote cache
●
IBM Power 5 continuous data sampling
●
X
2
5
Thread Clustering

10. Sharing Detection
To observe remote cache accesses:
●
Exploit HPCs (hardware performance counters)
●
Sample remote cache miss addresses
●
Local cache misses satisfied by remote cache
●
IBM Power 5 continuous data sampling
●
3
5
Thread Clustering

11. Sharing Signatures
Construct for each thread
●
Counts remote cache accesses
●
Conceptually
virtual address
virtual address
264
0
block
8-bit counter
6
Thread Clustering

12. Sharing Signatures
Construct for each thread
●
Counts remote cache accesses
●
Conceptually
virtual address
virtual address
264
0
ctri++
block
8-bit counter
6
Thread Clustering

13. Optimizations
CPU: Temporal Sampling
●
Sample every Nth remote cache access
●
Memory: Spatial Sampling
●
256-entry vector
●
Hash function
●
Block ID filter
●
Vectors still effective at indicating sharing
●
7
Thread Clustering

14. Spatial Sampling
Hash collision & alias removal
●
Filter Legend
Empty
Reserved
Block ID
255
0
255
0
8
Thread Clustering

15. Spatial Sampling
Hash collision & alias removal
●
Filter Legend
Empty
Reserved
hash
EMPTY
Block ID
255
0
255
0
8
Thread Clustering

16. Spatial Sampling
Hash collision & alias removal
●
Filter Legend
Empty
Reserved
hash
(First-Come-First-Reserved)
Block ID
255
0
255
0
8
Thread Clustering

17. Spatial Sampling
Hash collision & alias removal
●
Filter Legend
Empty
Reserved
hash
MATCH Block ID
Block ID
255
0
255
0
8
Thread Clustering

18. Spatial Sampling
Hash collision & alias removal
●
Filter Legend
Empty
Reserved
hash
MISMATCH Block ID
Block ID
255
0
ALIASING PREVENTED
255
0
8
Thread Clustering

19. Automated Clustering
Clustering Heuristic:
Simple, one-pass algorithm
●
Compare vector against existing clusters
●
If not similar, create a new cluster
●
Similarity Metric:
N
∑ V1[i] * V2[i]
i=0
Shared blocks amplified
●
Non-shared blocks nullified
●
9
Thread Clustering

20. Experimental Platform
8-way Power 5, 1.5GHz
●
Linux 2.6
●
IBM J2SE 5.0 JVM
●
1.9MB L2
1.9MB L2
36MB
36MB
4 GB 4 GB
10
Thread Clustering

21. Workloads
Microbenchmark
expect 4 clusters
●
4 threads per cluster
●
SPECjbb2000 (modified)
expect 2 clusters
●
2 warehouses, 8 threads per warehouse
●
RUBiS + MySQL
expect 2 clusters
●
2 databases, 16 threads per database
●
VolanoMark chat server
expect 2 clusters
●
2 rooms, 8 threads per room
●
11
Thread Clustering

22. Visualizing Clusters
Counter Values
An example 255
●
128
64
0
{
Cluster A,
4 vectors
{
Cluster B,
4 vectors
12
Thread Clustering

23. Visualizing Clusters
Counter Values
An example 255
●
128
64
0
{
Cluster A,
4 vectors
{
Cluster B,
4 vectors
12
Thread Clustering

24. Visualizing Clusters
Counter Values
An example 255
●
128
64
0
{
Cluster A,
4 vectors
{
Cluster B,
4 vectors
12
Thread Clustering

25. Visualizing Clusters
Microbenchmark
●
{
4
vectors
13
Thread Clustering

26. Visualizing Clusters
Modified SPECjbb2000 (4 warehouses)
●
{
16
vectors
14
Thread Clustering

27. Visualizing Clusters
RUBiS + MySQL (2 databases)
●
{
24
vectors
15
Thread Clustering

28. Visualizing Clusters
VolanoMark (4 rooms)
●
16
Thread Clustering

29. Remote Cache Impact
Normalized to default Linux
●
90
70 72
43
32
22
9
2
-17
17
Thread Clustering

30. Performance Impact
IPC: instructions per cycle
●
Normalized to default Linux
●
7.4 7.4
7.1
6.1 6.1
5.1
5.0
3.7
-0.8
18
Thread Clustering

31. Summary
AFTER:
BEFORE:
Operating System With
Current Operating Systems
Thread Clustering
19
Thread Clustering

32. Conclusions
HPCs can detect sharing
●
Sharing signatures are effective
●
Automated thread clustering:
●
Reduces remote cache access up to 70%
●
Improves performance up to 7%
●
All with low overhead
●
Future Work:
More workloads
●
Improve clustering algorithm
●
Integration with load-balancing aspects
●
20
Thread Clustering

33. Thread Clustering

34. Sampling Overhead
Modified SPECjbb2000
●
Thread Clustering