In modern Commercial Off-The-Shelf (COTS) multicore systems, each core can generate many parallel memory requests at a time. The processing of these parallel requests in the DRAM controller greatly affects the memory interference delay experienced by running tasks on the platform. In this paper, we present a new parallelism-aware worst-case memory interference delay analysis for COTS multicore systems. The analysis considers a COTS processor that can generate multiple outstanding requests and a COTS DRAM controller that has a separate read and write request buffer, prioritizes reads over writes, and supports out-of-order request processing. Focusing on LLC and DRAM bank partitioned systems, our analysis computes worst-case upper bounds on memory-interference delays, caused by competing memory requests. We validate our analysis on a Gem5 full-system simulator modeling a realistic COTS multicore platform, with a set of carefully designed synthetic benchmarks as well as SPEC2006 benchmarks. The evaluation results show that our analysis produces safe upper bounds in all tested benchmarks, while the current state-of-the-art analysis significantly under-estimates the delays.

1. Parallelism-Aware Memory Interference
Delay Analysis for COTS Multicore Systems
Heechul Yun +, Rodolfo Pellizzoni*, Prathap Kumar Valsan+
+University of Kansas
*University of Waterloo
1

2. High-Performance Multicores for
Embedded Real-Time Systems
• Why?
– Intelligence  more performance
– Space, weight, power (SWaP), cost
2

3. Challenge: Shared Memory Hierarchy
3
Core1 Core2 Core3 Core4
Memory Controller (MC)
Shared Cache
• Hardware resources are contented among the cores
• Tasks can suffer significant memory interference
delays
DRAM

4. Memory Interference Delay
• Can be extremely high in the worst-case
4
8.0
33.5
45.8
0
5
10
15
20
25
30
35
40
45
50
ARM
Cortex A15
Intel
Nahelem
Intel
Haswell
solo
+ 1 co-runner
+ 2 co-runners
+ 3 co-runners
Normalizedexecutiontime
45.8X
slowdown
DRAM
LLC
Core1 Core2 Core3 Core4
bench co-runner(s)
banks

5. Modeling Memory Interference
• Common (false) assumptions on COTS systems
– A single resource
• Reality: multiple parallel resources (banks)
– A constant memory service cost
• Reality: it varies depending on the DRAM bank state
– Round-robin arbitration, in-order processing
• Reality: FR-FCFS can re-order requests
– Both read and write requests are treated equally
• Reality: writes are buffered and processed opportunistically
– One outstanding request per core
• Reality: an out-of-order core can generate parallel reqs.
5
Addressed in This Work
Addressed in [Kim’14]
[Kim’14] H. Kim, D. de Niz, B. Andersson, M. Klein, O. Mutlu, and R. R.Rajkumar. “Bounding Memory Interference
Delay in COTS-based Multi-Core Systems,” RTAS’14

6. Our Approach
• Realistic memory interference model for COTS systems
– Memory-level parallelism (MLP) in COTS architecture
– Write-buffering and opportunistic batch processing in MC
• DRAM bank-partitioning
– Reduce interference
• Delay analysis
– Compute worst-case memory interference delay of the
task under analysis
6

7. Outline
• Motivation
• Background
– COTS multicore architecture
– DRAM organization
– Memory controller
• Our approach
• Evaluation
• Conclusion
7

8. Memory-Level Parallelism (MLP)
• Broadly defined as the number of concurrent
memory requests that a given architecture
can handle at a time
8

9. Last Level Cache (LLC)
DRAM DIMM
Memory Controller (MC)
Core1 Core2 Core3 Core4
Request buffers
Read Write
Scheduler
MSHRs
CMD/ADDR DATA
COTS Multicore Architecture
Out-of-order core:
Multiple memory requests
Non-blocking caches:
Multiple cache-misses
MC and DRAM:
Multiple banks
Bank
4
Bank
3
Bank
2
Bank
1
MSHRs MSHRs MSHRs MSHRs
9

10. DRAM Organization
LLC
DRAM DIMM
Memory Controller (MC)
Bank
4
Bank
3
Bank
2
Bank
1
Core1 Core2 Core3 Core4
Mess
• intra-bank conflicts
• Inter-bank conflicts
10

11. DRAM Bank Partitioning
L3
DRAM DIMM
Memory Controller (MC)
Bank
4
Bank
3
Bank
2
Bank
1
Core1 Core2 Core3 Core4 • Private banking
– OS kernel allocates
pages from
dedicated banks for
each core
Eliminate
intra bank
conflicts
11

12. Bank Access Cost
Row 1
Row 2
Row 3
Row 4
Row 5
A DRAM Bank
Row Bufferactivate
precharge
Read/write
• State dependent access cost
– Row hit: fast
– Row miss: slow
READ (Bank 1, Row 3, Col 7)
Col7
12

13. Memory Controller
13
Read request
buffer
Write request
buffer
Bank 1
scheduler
Channel scheduler
Bank 2
scheduler
Bank N
scheduler
DRAM chip
Memory requests from cores
Writes are buffered and processed opportunistically

14. “Intelligent” Read/Write Switching
• Intuition
– Writes are not in the critical path. So buffer and
process them opportunistically
• Algorithm [Hansson’14]
– If there are reads, process them unless the write
buffer is almost full (high watermark)
– If there’s no reads and there is enough buffered
writes (low watermark), process the writes until
reads arrive
14
[Hansson’14] Hansson et al., “Simulating DRAM controllers for future syste
m architecture exploration,” ISPASS’14

15. FR-FCFS Scheduling[Rixner’00]
• Priority order
1. Row hit request
2. Older request
• Maximize memory throughput
15
Bank 1
scheduler
Channel scheduler
Bank 2
scheduler
Bank N
scheduler
[Rixner’00] S. Rixner, W. J. Dally, U. J. Kapasi, P. Mattson, and J. Owens. Memo
ry access scheduling. ACM SIGARCH Computer Architecture News. 2000

16. Outline
• Motivation
• Background
• Our approach
– System model
– Delay analysis
• Evaluation
• Conclusion
16

17. System Model
• Task
– Solo execution time C
– Memory demand (#of LLC misses): H
• Core
– Can generate multiple, but bounded, parallel requests
• Upper-bounded by L1 cache’s MSHR size
• Cache (LLC)
– Assume no cache-level interference
• Core-private or partitioned LLC
• No MSHR contention
• DRAM controller
– Efficient FR-FCFS scheduler, open-page pollicy
– Separate read and write request buffer
• Watermark scheme on processing writes
17

18. Delay Analysis
• Goal
– Compute the worst-case memory interference delay of a
task under analysis
• Request driven analysis
– Based on the task’s own memory demand: H
– Compute worst-case per request delay: RD
– Memory interference delay = RD x H
• Job driven analysis
– Based on the other tasks’ memory requests over time
– See paper
18

19. Key Intuition #1
• The #of competing requests Nrq is bounded
– Because the # of per-core parallel requests is bounded.
– Example
• Cortex-A15’s per-core bound = 6
• Nrq = 6 x 3 (cores) = 18
19

20. Key Intuition #2
• DRAM sub-commands of the competing
memory requests are overlapped
– Much less pessimistic than [Kim’14], which simply
sums up each sub-command’s maximum delay
– See paper for the proof
20

21. Key Intuition #3
• The worst-case delay happens when
– The read buffer has Nrq requests
– And the write request buffer just becomes full
• Start a write batch
– Then the read request under analysis arrives
21
RD = read batch delay + write batch delay

22. Outline
• Motivation
• Background
• Our approach
• Evaluation
• Conclusion
22

23. Evaluation Setup
• Gem5 simulator
– 4 out-of-order cores (based on Cortex-A15)
• L2 MSHR size is increased to eliminate MSHR contention
– DRAM controller model [Hansson’14]
– LPDDR2 @ 533Mhz
• Linux 3.14
– Use PALLOC[Yun’14] to partition
DRAM banks and LLC
• Workload
– Subject: Latency, SPEC2006
– Co-runners: Bandwidth (write)
23
DRAM
LLC
Core1 Core2 Core3 Core4
subject co-runner(s)

24. Results with the Latency benchmark
• Ours(ideal): Read only delay analysis (ignore writes)
• Ours(opt): assume writes are balanced over multiple banks
• Ours(worst): all writes are targeting one bank & all row misses
24
0
5
10
15
20
25
30
Measured [Kim'14] Ours(ideal) Ours(opt) Ours(worst)
NormalizedResponseTime
underestimate
overestimate
[Kim’14] H. Kim, D. de Niz, B. Andersson, M. Klein, O. Mutlu, and R. R.Rajkumar. “Bounding Memory Interference
Delay in COTS-based Multi-Core Systems,” RTAS’14

25. Results with SPEC2006 Benchmarks
• Main source of pessimism:
– The pathological case of write (LLC write-backs) processing
25

26. Conclusion
• Memory interference delay on COTS multicore
– Existing analysis methods rely on strong assumptions
• Our approach
– A realistic model of COTS memory system
• Parallel memory requests
• Read prioritization and opportunistic write processing
– Request and job-driven delay analysis methods
• Pessimistic but still can be useful for low memory intensive tasks
• Future work
– Reduce pessimism in the analysis
26

27. Thank You
27