Post on 29-Sep-2020
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 1
UNIVERSITY OF MARYLAND
The Eff ects of OoO Ex ecution on The Memor y System
Ph.D. Research Proposal
Aamer Jaleelajaleel@eng.umd.edu
Proposal Committee:
Dr. Rajeev Barua {barua@eng.umd.edu}
Dr. Bruce Jacob {blj@eng.umd.edu}
Dr. Donald Yeung {yeung@eng.umd.edu}
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 2
UNIVERSITY OF MARYLAND
Overview
•
Motiv ation
•
The Problem
•
Effects on Repla y Traps
•
Effects on Cac he Performance
•
A New Metric -
Disorder
•
Absolute Disor der
•
Relative Disor der
•
Proposed Work
•
Sensitivity Studies
•
Dynamic Mec hanisms
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 3
UNIVERSITY OF MARYLAND
Out-of-Or der Ex ecution
•
Overlap idle time caused b y long latenc y operations with
possible
useful w ork
•
Floating point latenc y: 10-15 cycles
•
Memor y latenc y: 100-2000 cycles
•
Widel y held belief that a pr ocessor’ s OoO effi cienc y depends on the n umber of instructions it vie ws at a given time
Large Instruction Windo ws / ROBs!!!
More Instruction Level Parallelism!!!!!!!!
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 4
UNIVERSITY OF MARYLAND
OoO Hardware Structures
•
ROB, Integ er and Floating P oint Issue Queues, Load and Store Queues
ROBIQ
FQ
LQ
SQBP
IC
LP
RN
FETCHRENAME
UNIT
SCH
HD HD
HD
HD HD
TL
TL
TL
BP = Branch PredictorLP = Line Predictor
IC = Instruction CacheRN = Register Rename
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 5
UNIVERSITY OF MARYLAND
Research Trend in Increasing ILP
•
Proposals to b uild lar ge ROBs
•
Akkar y et al., Lebec k et al., Skadr on et al.
•
Circuit tec hniques to b uild lar ge ROBs with out aff ecting c loc k cycle time
•
Brown et al., Henry et al., Onder et al.
•
Techniques to allo w for more load/store queue comm unications
•
Park et al., Akkar y et al.
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 6
UNIVERSITY OF MARYLAND
Large Instruction Windo ws
•
Lebec k et al. sho w 35-250% perf ormance impr ovements with lar ge ROBs
•
However, when one consider s the discounted real eff ects , most of these perf ormance impr ovements disappear
•
Replay Traps
•
Cache Misses
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 7
UNIVERSITY OF MARYLAND
Replay Traps - Bac kgr ound
•
Replay traps are required to
•
Force accesses to a par ticular memor y location in or der
•
Handle diff erent-siz ed accesses to the same memor y location
•
Two reco very sc hemes
•
Flush pipeline , and then re-f etch and re-execute all instructions fr om the repla y trap causing instruction
•
No fl ush, re-execute ONLY repla y trap causing instruction and all other instructions that directl y or indirectl y depend on it
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 8
UNIVERSITY OF MARYLAND
Types of Repla y Traps
•
Load-Store Repla y
(a) Load-Store Repla y
2.
ST BYTE A (3)
3.
LD BYTE A (2)
1.
LD BYTE A (1)
4.
LD BYTE B (4)
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 9
UNIVERSITY OF MARYLAND
Types of Repla y Traps
•
Load-Store Repla y
•
Wrong Siz e Replay
(b) Wrong Siz e Replay
(a) Load-Store Repla y
2.
ST BYTE A (3)
3.
LD BYTE A (2)
1.
LD BYTE A (1)
4.
LD BYTE B (4)
2.
ST BYTE A (2)
3.
LD HALF A (3)
1.
LD BYTE A (1)
4.
LD BYTE B (4)
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 10
UNIVERSITY OF MARYLAND
Types of Repla y Traps
•
Load-Store Repla y
•
Wrong Siz e Replay
•
Multi-Pr ocessor
•
Load-Load Repla y
(b) Wrong Siz e Replay
(a) Load-Store Repla y
2.
ST BYTE A (3)
3.
LD BYTE A (2)
1.
LD BYTE A (1)
4.
LD BYTE B (4)
2.
ST BYTE A (2)
3.
LD HALF A (3)
1.
LD BYTE A (1)
4.
LD BYTE B (4)
2.
ST BYTE A (2)
3.
LD BYTE A (1)
1.
LD BYTE A (4)
4.
LD BYTE B (3)
P2P1
(c) Load-Load Repla y
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 11
UNIVERSITY OF MARYLAND
Types of Repla y Traps
•
Load-Store Repla y
•
Wrong Siz e Replay
•
Multi-Pr ocessor
•
Load-Load Repla y
•
Load-Miss Load Repla y
(b) Wrong Siz e Replay
(d) Load-Miss Load Repla y
(a) Load-Store Repla y
2.
ST BYTE A (3)
3.
LD BYTE A (2)
1.
LD BYTE A (1)
4.
LD BYTE B (4)
2.
ST BYTE A (2)
3.
LD HALF A (3)
1.
LD BYTE A (1)
4.
LD BYTE B (4)
2.
ST BYTE A (2)
3.
LD BYTE A (1)
1.
LD BYTE A (4)
4.
LD BYTE B (3)
P2P1
(c) Load-Load Repla y
3.
LD BYTE A (3)
2.
ST BYTE A (2)
1.
+
LD BYTE A (1)
4.
LD BYTE B (4)
P2P1
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 12
UNIVERSITY OF MARYLAND
The Problem
•
Aggressive OoO tec hniques result in an
•
Increase in frequenc y of
replay traps
•
Increase in n umber of L1
cache misses
appluartmgridswimgcc
gzipmcf
twolfFAvg
IAvg0
20
40
60
appluartmgridswimgcc
gzipmcf
twolfFAvg
IAvg0
10
20
30
40
% Total Ex ecution Time Spent in Handling Traps % Increase in L1 Cac he Misses Compared to 80-Entr y ROB
80 Entry Rob128 Entry Rob256 Entry Rob512 Entry Rob
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 13
UNIVERSITY OF MARYLAND
Hypothesis
•
Increasing R OB siz es reor der ALU and memor y instructions
•
Re-ordering of ALU instructions poses little or no threats, BUT re-or dering of memor y instructions causes most of the negative eff ects
•
3 Processor Confi gurations
•
ALU-in / MEM-in : ALU and memor y instructions issued in-or der (
sequential
)
•
ALU-out / MEM-in : ALU instructions issued out-of-or der, memor y instructions issued in-order (
more speculation than ALU-in/MEM-in
)
•
ALU-out / MEM-out : Both ALU and memor y instructions are issued OoO (
superscalar
)
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 14
UNIVERSITY OF MARYLAND
Experimental Frame work
•
Sim-Alpha
•
Validated Ex ecution Driven Sim ulator
•
Detailed DRAM Sim ulator
•
64K 2-way IL1/DL1 and 2MB 4-wa y Unifi ed L2
•
8 MSHRS per cac he
•
1024-entr y store-wait data structure
•
Suppor ts No/ Sequential /Stride Data Pref etch
•
Benc hmarks
•
SPEC2000 & Olden
Table 1: Processor Parameters
Configuration Name
ROB Size Issue WidthIssueQ Size
INT/FP# Functional
Units
**
LQ/SQSize
Alpha-80 80 2-32 Way 20/15 (4/4/1/1) x1 32/32Alpha-128 128 2-32 Way 40/30 (4/4/1/1) x2 64/64Alpha-256 256 2-32 Way 80/60 (4/4/1/1) x4 128/128Alpha-512 512 2-32 Way 160/120 (4/4/1/1) x8 256/256
**INT ALU/INT MULT/FP ALU/FP MULT
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 15
UNIVERSITY OF MARYLAND
OoO And Repla y Trap Overhead
•
ALU-in/MEM-in --> ALU-out/MEM-out
•
Factor 8-15 increase in trap frequenc y
•
15-30% increase in trap o verhead
8-Way 808-Way 128
8-Way 2568-Way 512
0
10
20
30
40
50
Trap
Ove
rhea
d: C
ycle
s Lo
st in
Tra
ps
ALU-in / MEM-inALU-out / MEM-inALU-out / MEM-out
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 16
UNIVERSITY OF MARYLAND
OoO And Cac he Misses
•
ALU-in/MEM-in --> ALU-out/MEM-out
•
25-75% increase in L1 Cac he Misses
•
80-125% increase in L2 Cac he Misses
8-Way 808-Way 128
8-Way 2568-Way 512
0
25
50
75
100
8-Way 808-Way 128
8-Way 2568-Way 512
0
25
50
75
100
125
150
175
200
% Increase in L2 Cache Misses% Increase in L1 Cache Misses
Normalized to # L1 Cache Misses of ALU-in/MEM-in
ALU-out / MEM-inALU-out / MEM-out
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 17
UNIVERSITY OF MARYLAND
What Have We Learned?
•
In general, increasing a pr ocessor’ s OoO capability confl icts with the memor y ordering requirements and cac he perf ormance
•
Limiting the memor y instructions to be issued in pr ogram or der aids in reducing these negative eff ects
•
Downside of MEM in-or der: memor y ILP hur t
•
Future a ggressive OoO pr ocessor s need a mechanism to thr ottle the degree b y whic h the y issue memor y instructions out-of-or der
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 18
UNIVERSITY OF MARYLAND
Intr oducing A Ne w Metric
•
Disorder
— The degree b y whic h an instruction is issued out-of-or der
•
Two types of disor der
•
Absolute disor der — On a pr ogram or der perspective
•
Relative disor der — On a per instruction executed per spective
INST 1
INST 2
INST 3
INST 4
INST 5
INST 6
INST 7
INST 6
INST 3
INST 1
INST 4
INST 5
INST 7
INST 2
Absolute Disor der
Initial Pr ogram Or der Executed Pr ogram Or der
Relative Disor der
Scheduler
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 19
UNIVERSITY OF MARYLAND
Disor der Measurements
•
We measure disor der ONLY for memor y instructions
•
When a memor y instruction is f etched, it is assigned a sequential ID , i.e. the fi rst memor y instruction f etched g ets sequential ID 1, next g ets sequential ID 2, and so on.
•
In the e vent of a pipeline fl ush, the sequential ID is restored to the last successfull y retired sequential ID + 1
•
Disor der is computed after a memor y instruction has its dependencies resolved and is issued to e xecute
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 20
UNIVERSITY OF MARYLAND
Absolute Disor der
•
Degree b y whic h an instruction is issued OoO compared to f etch or der
•
Computed b y fi nding the diff erence between the instruction issued and the instruction that
should’ve been issued
were the core in-or der
MEM
1
MEM
2
MEM
3
MEM
4
MEM
5
MEM
6
MEM
7
MEM
8
MEM
9
MEM
10
MEM
11
MEM
1
MEM
3
MEM
5
MEM
7
MEM
8
MEM
2
MEM
10
MEM
4
MEM
9
MEM
11
MEM
6
0
1
2
3
3
- 4
3
- 4
0
1
- 5
4 - Way Issue Order
Cycle 101: 1, 3Cycle 105: 5, 7, 8Cycle 126: 2, 10Cycle 139: 4Cycle 213: 9, 11Cycle 224: 6
PROGRAM ORDER ISSUE ORDER ABSOLUTE DISORDER
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 21
UNIVERSITY OF MARYLAND
Relative Disor der
•
Degree b y whic h an instruction is issued OoO compared to other instructions
•
Computed b y fi nding the minim um absolute disor der between an instruction and all other instructions issued in the current and a pre vious c ycle
MEM
1
MEM
2
MEM
3
MEM
4
MEM
5
MEM
6
MEM
7
MEM
8
MEM
9
MEM
10
MEM
11
1
-3
2
2
2
3
-1
1
-2
2
2
4 - Way Issue Order
Cycle 101: 1, 3Cycle 105: 5, 7, 8Cycle 126: 2, 10Cycle 139: 4Cycle 213: 9, 11Cycle 224: 6
Relative Disorder of
10 - 2 = 810 - 5 = 510 - 7 = 310 - 8 = 2
MEM10
Minimum = 2
RELATIVE DISORDER
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 22
UNIVERSITY OF MARYLAND
Disor der Trends
•
Vary Issue Widths fr om 4 to 32 wide
•
Vary ROB Sizes fr om 80 to 512 entries
Increasing Issue WidthsIn
crea
sing
RO
B S
izes
Absolute Disor derRelative Disor der
+ Disor der
%In
str
60%
30%
70%
9%
- Disor der
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 23
UNIVERSITY OF MARYLAND
Disor der Results
•
Low disor der is due to dependenc y stalls and L1 miss penalty . High disor der is due to L2 miss penalty
•
Absolute Disor der
•
Increasing R OBs and not issue widths cause an increase in absolute disor der
•
< 1/3
rd
of TOTAL memor y instructions executed are issued in actual pr ogram or der
•
Relative Disor der
•
60 – 70% of memor y instructions issued are in c lose pr oximity with eac h other in the instruction stream
•
Spatial issue among memor y instructions, i.e. “when a pr ocessor speculates it contin ues to speculate”
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 24
UNIVERSITY OF MARYLAND
Proposed Work
•
Sensitivity Studies
•
Cache Organization P arameter s
•
Load/Store Queue Studies
•
Dynamic Mec hanisms
•
Windo wing of Load/Store Queue
•
?????
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 25
UNIVERSITY OF MARYLAND
Disor der Vs. Cache Organization
•
Cache Parameter s
•
Cache Size
•
Cache Line Siz e
•
Cache Associativity
Vary Cache Associativity
Vary
Cach
e Size
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 26
UNIVERSITY OF MARYLAND
Disor der Vs. Load/Store Queue
•
Intr oduce a vir tual windo w into the load/store queue
•
Only instructions residing within the vir tual windo w may be issued
•
Other s wait until the vir tual windo w slides onto them
LSQ Tail
LSQ Head
LD/ST 1LD/ST 2
LD/ST 4
.LD/ST 5
LD/ST 0
LD/ST 3
.LD/ST N-1LD/ST N
Vir tualWindo wSize=Inf
Traditional Load/Store Queue
LD/ST 1LD/ST 2
LD/ST 4
.LD/ST 5
LD/ST 0
LD/ST 3
Vir tual Head
Vir tual Tail
.LD/ST N-1LD/ST N
LSQ Tail
LSQ Head
Vir tualWindo wSize=4
Vir tual Load/Store Queue
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 27
UNIVERSITY OF MARYLAND
Windo wing of Load/Store Queue
•
Static Mec hanism (Sensitivity Stud y):
•
Staticall y set the siz e of the vir tual windo w based on pr ofi le inf ormation of the application in concern
•
Drawbac k: Memor y ILP is lost during periods of application e xecution where negative effects do not e xist
•
Dynamic Mec hanism:
•
Dynamicall y vary the siz e of the vir tual windo w based on application beha vior during execution
•
Vir tual windo w initiall y star ts at infi nity b ut as cer tain thresholds are reac hed the vir tual windo w siz e is reduced
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 28
UNIVERSITY OF MARYLAND
Conc lusions
•
Though the well kno wn mec hanism of increasing ILP impr oves perf ormance , it can cause side eff ects in the memor y system
•
Characteriz ed the pr oblem in terms of
•
Increased Repla y Traps
•
Increased Cac he Misses
•
Sour ce of pr oblem is the reor dering of memor y instructions
•
Proposing to stud y mec hanisms to thr ottle the degree b y whic h memor y instructions are issued OoO
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 29
UNIVERSITY OF MARYLAND
Questions?????
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 30
UNIVERSITY OF MARYLAND
Impr oving P erformance
Execution Time = Cyc le time * CPI * Inst Cnt
•
Instruction Le vel Parallelism (ILP)
•
Pipelining
•
Multiple Issue Width
•
Out-of-Or der Ex ecution
•
Speculation
•
Cache Line Prediction
•
Branc h Prediction
•
Prefetching
•
Hardware Pref etching
•
Software Pref etching
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 31
UNIVERSITY OF MARYLAND
Industr y Trends
•
Two Design Philosophies
•
Brainiacs
: Impr ove micr opr ocessor perf ormance b y increasing ILP
•
Speed Demons
: Impr ove micr opr ocessor perf ormance b y increasing c loc k speeds
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 32
UNIVERSITY OF MARYLAND
Why Use a Vir tual Windo w?
•
Provide a mec hanism to allo w for lar ge ROBs to e xploit ALU instruction ILP y et provide the benefi ts of smaller R OBs
•
Since memor y instructions held in load/store queues (LSQ), could ha ve lar ge ROB siz es and small LSQs.
•
Effective in limiting n umber of memor y instructions in fl ight, hence reduces disor der
•
Ineffi cient design methodology as it can under utiliz e the ROB space
•
Allo w for lar ge ROB and LSQ siz es but create a
virtual window
in LSQ
•
Staticall y or d ynamicall y vary the siz e of vir tual windo w
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 33
UNIVERSITY OF MARYLAND
Performance Graphs
art gccmcf
parserperlbmk
swimtwolf
vortex vpr
0
1
2
3
(ALU-out / MEM-out)
art gccmcf
parserperlbmk
swimtwolf
vortex vpr
0
1
2
3
art gccmcf
parserperlbmk
swimtwolf
vortex vpr
0
1
2
3
(ALU-in / MEM-in)
(ALU-out / MEM-in)
MemoryALUOverhead
abcdefghi
a - 2-Way ROB 80b - 4-Way ROB 80c - 8-Way ROB 80d - 4-Way ROB 128
e - 8-Way ROB 128f - 4-Way ROB 256g - 8-Way ROB 256h - 4-Way ROB 512i - 8-Way ROB 512
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 34
UNIVERSITY OF MARYLAND
Performance Graphs
art gccmcf
parserperlbmk
swimtwolf
vortex vprbisort
em3dhealth mst
perimetertreeadd
0
1
2
3
4
5
6
7
art gccmcf
parserperlbmk
swimtwolf
vortex vprbisort
em3dhealth mst
perimetertreeadd
0
1
2
3
4
5
6
7
art gccmcf
parserperlbmk
swimtwolf
vortex vprbisort
em3dhealth mst
perimetertreeadd
0
1
2
3
4
5
6
7
(ALU-in / MEM-in)
(ALU-out / MEM-in)
(ALU-out / MEM-out)
Memory Stall in CommitNon-Memory Stall in CommitOverhead
abcdefghi
a - 2-Way ROB 80b - 4-Way ROB 80c - 8-Way ROB 80d - 4-Way ROB 128e - 8-Way ROB 128f - 4-Way ROB 256g - 8-Way ROB 256h - 4-Way ROB 512i - 8-Way ROB 512
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 35
UNIVERSITY OF MARYLAND
The Problem - Cac he Misses
•
Aggressive OoO tec hniques result in an increase in application
cache misses
ammpapplu
apsi artequake
fma3dgalgel
lucasmesa
mgridswimwupwise
bzip2craftyeon gap gcc
gzipmcf
parser
perlbmktwolf
vortexvpr
FAVGIAVG
0
10
20
30
40
80 Entry Rob128 Entry Rob256 Entry Rob512 Entry Rob
% Increase in L1 Cac he Misses Compared to 80-Entr y ROB
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 36
UNIVERSITY OF MARYLAND
The Problem - Repla y Traps
•
Aggressive OoO tec hniques result in an increase in
replay traps
•
Aggravate with increasing R OB siz es.
80 Entry Rob128 Entry Rob256 Entry Rob512 Entry Rob
ammpapplu
apsi artequake
fma3dgalgel
lucasmesa
mgridswimwupwise
bzip2craftyeon gap gcc
gzipmcf
parser
perlbmktwolf
vortexvpr
FAVGIAVG
0
20
40
60
80
% Total Ex ecution Time Spent in Handling Traps
PhD Research
Proposal
The Effects Of OoOExecution On The
Memory System
Aamer Jaleel
University ofMaryland
ECE Dept.
SLIDE 37
UNIVERSITY OF MARYLAND
•
2-Way 808-Way 80
8-Way 1288-Way 256
8-Way 512-30
-25
-20
-15
-10
-5
0
Decrease in L1 Cache Misses
2-Way 808-Way 80
8-Way 1288-Way 256
8-Way 512
1
1e1
1e2
1e3
1e4
1e5
1e6
Trap Frequency: Instructions/Trap