Datacenter application interference CMPs (popular in datacenters) - - PowerPoint PPT Presentation

▶

Apr 27, 2023 198 likes •561 views

Datacenter application interference CMPs (popular in datacenters) offer increased throughput and reduced power consumption They also increase resource sharing between applications, which can result in negative interference. 1 Resource

SLIDE 1

Datacenter application interference

CMPs (popular in datacenters) offer increased

throughput and reduced power consumption

They also increase resource sharing between

applications, which can result in negative interference.

SLIDE 2

Resource contention is well studied … at least on single machines. 3 main methods: (1) Gladiator style match-ups (2) Static analysis to predict application resource usage (3) Measure benchmark resource usage; apply to live applications

SLIDE 3

New methodology for understanding datacenter interference is needed.

One that can handle complexities of a datacenter:  (10s of) thousands of applications  real user inputs  production hardware  financially feasible  low overhead Hardware counter measurements of live applications.

SLIDE 4

Our contributions

1. ID complexities in datacenters
2. New measurement methodology
3. First large-scale study of measured

interference on live datacenter applications.

SLIDE 5

Complexities of understanding application interference in a datacenter

SLIDE 6

Large chips and high core utilizations Profiling 1000 12-core,

24-hyperthread Google servers running production workloads revealed the average machine had >14/24 HW threads in use.

SLIDE 7

Heterogeneous application mixes

Often applications have more than

ne co-runner on a machine.

Observed max of 19 unique co- runner threads (out

f 24 HW

threads). 0-1 Co-runners 2-3 Co-runners 4+ Co-runners

SLIDE 8

Application complexities

Fuzzy definitions
Varying and sometimes

unpredictable inputs

Unknown optimal performance

SLIDE 9

Hardware & Economic Complexities

Varying micro-arch platforms
Necessity for low overhead =

limited measurement capabilities

Corporate policies

SLIDE 10

Measurement methodology

SLIDE 11

Measurement Methodology

The goal: A generic methodology to collect application interference data

n live production

datacenter servers

SLIDE 12

Measurement Methodology

App. A
App. B

Time

SLIDE 13

Measurement Methodology

1. Use sample-

based monitoring to collect per machine per core event (HW counter) sample data.

1.

SLIDE 14

Measurement Methodology

App. A
App. B

2 M instrs 2 M instrs 2 M instrs 2 M instrs 2 M instrs 2 M instrs 2 M instrs 2 M instrs 2 M instrs 2 M instrs

1 2 3 4 5 6 1 2 3 4

SLIDE 15

Measurement Methodology

2. Identify sample

sized co-runner relationships…

2.

SLIDE 16

Measurement Methodology

App. A
App. B

Samples A:1- A:6 are co-runners with App. B. Samples B:1- B:4 are co-runners with App. A.

SLIDE 17

Measurement Methodology

App. C
App. A
App. B

Say that a new App. C starts running

n CPU 1…

… B:4 no longer has a co-runner.

SLIDE 18

Measurement Methodology

3. Filter

relationships by

arch. independent

interference classes…

3.

SLIDE 19

Measurement Methodology

Be on opp. sockets.

SLIDE 20

Measurement Methodology

Share only I/O

SLIDE 21

Measurement Methodology

4. Aggregate

equivalent co- schedules.

4.

SLIDE 22

Measurement Methodology

For example:

Aggregate all the samples of App. A

that have App. B as a shared core co- runner.

Aggregate all samples of App. A that

have App. B as a shared core co-runner and App. C as a shared socket co- runner.

SLIDE 23

Measurement Methodology

5. Finally, calculate

statistical indicators (means, medians) to get a midpoint performance for

app. interference

comparisons

5.

SLIDE 24

Measurement Methodology

App. A
App. B
Avg. IPC

= 2.0

Avg. IPC

= 1.5

SLIDE 25

Applying the measurement methodology at Google.

SLIDE 26

Applying the Methodology @ Google

Event Instrs  IPC Sampling period 2.5 Million Number of machines* 1000

Experiment Details:

* All had Intel Westmere chips (24 hyperthreads, 12 cores), matching clock speed, RAM, O/S

1. Collect

samples

Method:

SLIDE 27

Applying the Methodology @ Google

Event Instrs  IPC Sampling period 2.5 Million Number of machines* 1000

Experiment Details:

* All had Intel Westmere chips (24 hyperthreads, 12 cores), matching clock speed, RAM, O/S Unique binary apps 1102 Co-runner relationships (top 8 apps)

Avg. shared core rel’ns 1M (min 2K)
Avg. shared socket

9.5M (min 12K)

Avg. opposite socket

11M (min 14K)

Collection results:

1. Collect

samples

Method:

2. ID sample

size relationships

3. Filter by

interference classes

SLIDE 28

Applying the Methodology @ Google

4. Aggregate

equiv. schedules

Method:

5. Calculate

statistical indicators

SLIDE 29

Analyze Interference

streeview’s IPC changes with top co-runners

Overall median IPC across 1102 applications

SLIDE 30

Beyond noisy interferers (shared core)

Co-running applications Base Application Less or pos. interference Negative interference Noisy data

SLIDE 31

Beyond noisy interferers (shared core)

* Recall minimum pair has 2K samples; medians across full grid of 1102 apps

Base Applications Co-running applications Less or pos. interference Noisy data Negative interference

SLIDE 32

Performance Strategies

Restrict negative beyond noisy

interferers (or encourage positive interferers as co-runners)

Isolate sensitive or antagonistic

applications

SLIDE 33

Takeaways

1. New datacenter application

interference studies can use our identified complexities as a check list.

2. Our measurement methodology

(verified at Google in 1st large-scale measurements of live datacenter interference), is generally applicable and shows promising initial performance opportunities.

SLIDE 34