Interference in Branch Predictors: A Systematic Approach

1
Interference in Branch Predictors A Systematic
Approach

• Veerle Desmet, Hans Vandierendonck, Koen De
  Bosschere
• Ghent University UGent
• ELIS/PARIS, member

PROBLEM INTERFERENCE
METHODOLOGY PROFILING CLUSTERING
State in branch predictors is often shared
between uncorrelated branches due to space
limitations
Branches with similar behavior can share
prediction information 1 1 1 1 0 0 0 0 1 1 1 1 0
1 0 1 1 1 1 1 0 0 0 0 1 1 1 1 0 1 0 Also valid
for branches that do not interleave in time,
e.g. 1 1 1 1 0 0 0 0 1 1 0 1 0
3 bit index
Index101
A
Index101
B
metric
prediction table
Time varying taken rate behavior of static
branches in time slices of 10M conditional
branch executions
destructive
50
constructive
45
neutral prediction identical to
unlimited predictor
40
Each branch represents a point in the N-dim taken
rate space
35
30
interference rate ()
25
20
K-MEANS CLUSTERING
15
10
1
2
5
0
X
16
32
64
128
256
512
1024
2048
4096
8192
16384
32768
65536
13107
26214
52428
Cluster
X
Cluster
Cluster
predictor size (bytes)
initial centers
calculate nearest center
Cluster
Cluster
Cluster
Cluster
Interfering pairs determined by
address Uncontrolled amount of interference System
atic approach?
X
Self-profiling versus cross-profiling
X
3
4
restart with new centers
redefine centers
RESULTS SYSTEMATICALLY DEFINED CLUSTERS IMPROVE
VARIOUS PREDICTION SCHEMES
SUBTABLES PER CLUSTER
HASHING
Original history length shorter branch address
gshare-like
applicable to every prediction scheme
destructive
3 bit index
3 bit index
50
constructive
45
History
Cluster
40
neutral prediction identical to
unlimited predictor
Cluster
PC
35
30
interference rate ()
Hashed index
25
20
Shorter original index
15
25
10
5
prediction table
prediction table
20
0
16
32
64
128
256
512
1024
2048
4096
8192
16384
32768
65536
13107
26214
52428
15
misprediction rate ()
predictor size (bytes)
10
5
0
Gshare predictor original cross
profiling self profiling Bi-mode Gskew
Perceptron predictor original cross
profiling self profiling
10
100
1000
12
Gshare predictor original cross
subtables self subtables cross hashing self
hashing
25
25
7
19
20
20
6,5
6
15
15
misprediction rate ()
misprediction rate ()
misprediction rate ()
5,5
10
10
5
5
4,5
5
5
4
0
0
3,5
100
1000
10000
100000
1000
10000
100000
1000000
10
100
1000
10000
100000
1000000
predictor size (bytes)
predictor size (bytes)
predictor size (bytes)