TOWARD PREDICTABLE PERFORMANCE IN SOFTWARE PACKET-PROCESSING PLATFORMS. Mihai Dobrescu, EPFL Katerina Argyraki, EPFL Sylvia Ratnasamy, UC Berkeley

Transcription

1 TOWARD PREDICTABLE PERFORMANCE IN SOFTWARE PACKET-PROCESSING PLATFORMS Mihai Dobrescu, EPFL Katerina Argyraki, EPFL Sylvia Ratnasamy, UC Berkeley

2 Programmable Networks 2 Industry/research community efforts Easily deploy new services Test research ideas Software packet processing General purpose hardware Familiar programming environment Extensible network functionality

3 Problem: Unpredictable Performance 3 Resource contention Caches, memory controllers, buses Performance interference Software packet-processing systems [Dobrescu 09, Han 10] High performance Same processing for all packets Goal: software packet processing with predictable performance

4 4 System Overview Contention for shared resources IP forward Filtering IP forward Encryption IP forward Statistics input traffic packet processing output traffic General purpose server

5 Is This Hard? 5 Yes, in general-purpose context Math models to predict contention Contention-aware job placement In packet-processing context?

6 Our Contribution 6 1. It is feasible to build a packet-processing platform with predictable performance using simple techniques. 2. Contention-aware job placement does not bring significant benefit to the overall performance.

7 Outline 7 System overview Contention factors Observations on application behavior A simple prediction method Intuition

8 System Overview 8 IP forward Filtering IP forward Encryption IP forward Statistics input traffic packet processing output traffic General purpose server

9 Workloads 9 Application Main functionality Characteristics IP IP routing, 128k entries L3 cache intensive MON Monitoring, 100k flows L3 cache intensive FW Firewall, 1000 rules L2 cache intensive RE Redundancy elimination Memory intensive VPN Encryption CPU intensive Synthetic Random cache reads Cache/memory/CPU Representative set of realistic applications

10 DRAM Memory Controller Setup 10 Linux + Click Commodity Intel Xeon server DRAM Shared L3 Cache Bus Shared L3 Cache Memory Controller

11 Basic Configuration 11 One application per core NUMA-aware memory allocation DRAM Memory Controller Shared L3 Cache Bus Shared L3 Cache Memory Controller DRAM Contention domain Contention domain Contended resources: cache and memory controller

12 Resource Contention Effects 12 Performance Drop (%) IP competitors 5 MON competitors 5 FW competitors 5 RE competitors 5 VPN competitors 0 IP MON FW RE VPN

13 Outline 13 System overview Contention factors Observations on application behavior A simple prediction method Intuition

14 Contention Factors 14 5 synthetic competitors Performance Drop (%) IP MON FW RE VPN 0 Cache + Memory Controller Contention Cache Contention Memory Controller Contention Cache is the dominant contention factor

15 Outline 15 System overview Contention factors Observations on application behavior A simple prediction method Intuition

16 Characterize Application Behavior continuous curves: synthetic competitors Performance Drop (%) Competitors L3 refs/sec (M)

17 Characterize Application Behavior continuous curves: synthetic competitors Performance Drop (%) individual points: realistic competitors Competitors L3 refs/sec (M)

18 Characterize Application Behavior continuous curves: synthetic competitors Performance Drop (%) individual points: realistic competitors Competitors L3 refs/sec (M) Obs. #1: competitors cache refs/sec determine drop

19 Characterize Application Behavior 19 Performance Drop (%) IP MON FW RE VPN Competitors L3 refs/sec (M) continuous curves: synthetic competitors individual points: realistic competitors Obs. #1: competitors cache refs/sec determine drop

20 Characterize Application Behavior 20 Performance Drop (%) IP MON FW RE VPN Competitors L3 refs/sec (M) continuous curves: synthetic competitors individual points: realistic competitors Obs. #2: drop curve grows slowly after certain point

21 Outline 21 System overview Contention factors Observations on application behavior A simple prediction method Intuition

22 Contention Effects Prediction 22 Step#1: performance drop curve for each app Synthetic competitors random cache reads Vary competitors cache refs/sec Step#2: cache refs/sec for each app running alone Step#3: predicted drop equals the value of the drop curve corresponding to the competing cache refs/sec Simple offline profiling

23 Step by Step Prediction Performance Drop (%) #1 Drop Curve #2 Competitors cache refs/sec Competitors L3 refs/sec (M) Simple offline profiling

24 Prediction Errors Performance Drop (%) Measured Drop #1 Drop Curve #2 Competitors cache refs/sec Competitors L3 refs/sec (M)

25 Evaluation 25 Error in Predicting Performance Drop IP MON FW RE VPN Contention effects are predictable 5 IP competitors 5 MON competitors 5 FW competitors 5 RE competitors 5 VPN competitors

26 Outline 26 System overview Contention factors Observations on application behavior A simple prediction method Intuition

27 The Intuition 27 Obs. #1: competitors cache refs/sec determine drop Aggregate data exceeds cache size 3MB shared cache/core

28 The Intuition 28 Obs. #1: competitors cache refs/sec determine drop Aggregate data exceeds cache size 3MB shared cache/core Obs. #2: drop curve grows slowly after certain point Most damage happens early on Simple probabilistic analysis

29 Conclusion 29 It is feasible to build a packet-processing platform with predictable performance using simple techniques 3% prediction error Contention-aware job placement does not bring significant benefit to the overall performance 2% potential improvement