TraceSplitter: A New Paradigm for Downscaling Traces

This study introduces TraceSplitter, a novel approach for downscaling traces to run on smaller setups effectively. It discusses the need for downscaling in production and experimental setups, focusing on load balancing techniques. The paper compares prior approaches such as timespan scaling, average rate scaling, and random sampling, highlighting the benefits of TraceSplitter in handling distorted temporal patterns and overloads. Key ideas include using load balancing techniques to optimize downscaling for improved system performance.

• Traces
• Downscaling
• Load Balancing
• Production Setup
• Experimental Setup

mial_2 Follow

Uploaded on Oct 10, 2024 | 0 Views

Download Presentation

Please find below an Image/Link to download the presentation.

The content on the website is provided AS IS for your information and personal use only. It may not be sold, licensed, or shared on other websites without obtaining consent from the author. Download presentation by click this link. If you encounter any issues during the download, it is possible that the publisher has removed the file from their server.

Presentation Transcript

1. TraceSplitter: A New Paradigm for Downscaling Traces Sultan Mahmud Sajal, Rubaba Hasan, Timothy Zhu, Bhuvan Urgaonkar, Siddhartha Sen* The Pennsylvania State University, Microsoft Research* 1

2. The Need for Downscaling Production Setup Experimental Setup In Memory Cache Web Server In Memory Cache Web Server User Traffic Database Database Load Too High arrival time, url, payload size 150382450, www.foo.com, 500 193720475, www.bar.com, 20 215820849, www.abc.com, 472 222941037, www.efg.com, 82 230491837, www.gka.com, 129 249104962, www.ndw.com, 901 ... 11 12 2

3. Downscaling: lower a trace's load to run on a smaller setup Production Setup Experimental Setup In Memory Cache Web Server In Memory Cache Web Server User Traffic Database Database Downscale arrival time, url, payload size 150382450, www.foo.com, 500 193720475, www.bar.com, 20 215820849, www.abc.com, 472 222941037, www.efg.com, 82 230491837, www.gka.com, 129 249104962, www.ndw.com, 901 ... 11 12 3

4. Prior Approaches: Timespan Scaling Original Trace OSDI 16, ICAC 16 System Capacity Load Distorted Temporal Pattern Longer Lasting Overloads Time Downscaled Trace Load System Capacity Time 4

5. IEEE NOMS 12, ICDCS 18 Prior Approaches : Average Rate Scaling Load Load Load Load Time Time Time Time Original Trace Downscaled Trace 5

6. Prior Approaches : Random Sampling SIGCOMM 19, IC2E 14 0.4 0.35 % of rare requests 0.3 0.25 0.2 time 0.15 0.1 . . . 0.05 0 Original Rand-1 Rand-2 Rand-3 Rand-4 Rand-5 Rand-6 Rand-7 Rand-8 Rand-9 Rand-10 Original Trace Downscaled Traces 6

7. Key Idea: Use Load Balancing Techniques for Trace Downscaling arrival time, url, payload size 150382450, www.foo.com, 500 arrival time, url, payload size 150382450, www.foo.com, 500 193720475, www.bar.com, 20 Production Setup In Memory Cache Web Server arrival time, url, payload size 193720475, www.bar.com, 20 arrival time, url, payload size 215820849, www.abc.com, 472 arrival time, url, payload size 215820849, www.abc.com, 472 222941037, www.efg.com, 82 User Traffic Database arrival time, url, payload size 222941037, www.efg.com, 82 arrival time, url, payload size 150382450, www.foo.com, 500 193720475, www.bar.com, 20 215820849, www.abc.com, 472 222941037, www.efg.com, 82 230491837, www.gka.com, 129 249104962, www.ndw.com, 901 arrival time, url, payload size 230491837, www.gka.com, 129 230491837, www.gka.com, 129 249104962, www.ndw.com, 901 arrival time, url, payload size arrival time, url, payload size 249104962, www.ndw.com, 901 ... 11 7

8. TraceSplitter Framework Scaling Factor, f arrival time, url, payload size 150382450, www.foo.com, 500 215820849, www.abc.com, 472 TraceSplitter arrival time, url, payload size 193720475, www.bar.com, 20 249104962, www.ndw.com, 901 arrival time, url, payload size 222941037, www.efg.com, 82 230491837, www.gka.com, 129 Downscaling Technique Random Sampling RR Randomized Round Robin (TS-RRR) TS-LWL Simulated Load Balancer arrival time, url, payload size 150382450, www.foo.com, 500 193720475, www.bar.com, 20 215820849, www.abc.com, 472 222941037, www.efg.com, 82 230491837, www.gka.com, 129 249104962, www.ndw.com, 901 Load Balancing Policy Random Round Robin Join Shortest Queue Least Work Left 8

9. TraceSplitter in Action Production Setup In Memory Cache Web Server User Traffic Experimental Setup Database In Memory Cache Scaling Factor, f Web Server TraceSplitter Database 11 time Simulated Load Balancer 12 13 Original Trace 9

10. TraceSplitter in Action Production Setup In Memory Cache Web Server User Traffic Experimental Setup Database In Memory Cache Scaling Factor, f Web Server TraceSplitter Database 11 time Simulated Load Balancer 12 13 Original Trace 10

11. TraceSplitter in Action Production Setup In Memory Cache Web Server User Traffic Experimental Setup Database In Memory Cache Scaling Factor, f Web Server TraceSplitter Database 11 time Simulated Load Balancer 12 13 Original Trace 11

12. Evaluation Methodology Production Setup Experimental Setup In Memory Cache Web Server In Memory Cache Web Server User Traffic Database Database arrival time, url, payload size 150382450, www.foo.com, 500 193720475, www.bar.com, 20 215820849, www.abc.com, 472 222941037, www.efg.com, 82 230491837, www.gka.com, 129 249104962, www.ndw.com, 901 ... metric metric Downscale Average Latency 50th Percentile Latency 90th Percentile Latency 99th Percentile Latency Average Latency 50th Percentile Latency 90th Percentile Latency 99th Percentile Latency Compare 12

13. Result from Microsoft OneRF Trace 7 Closer to Original is Better 6 Original TS-LWL TS-RRR RR AvgRate Rand T-span 5 Latency (s) 4 3 2 1 0 mean 50 pctl 90 pctl 95 pctl 99 pctl 13

14. Result from Synthetic Trace 11.56 13.05 18.03 6 5 4 Latency (s) 3 2 1 0 mean 50 pctl 90 pctl 95 pctl 99 pctl 14

15. Autoscaler Case Study Goal: 99th percentile latency (P99) < 500ms Define: Overprovisioning factor = c Objective: Learn correct policy to preserve latency goal Policy: cmin = minimum overprovisioning factor to meet goal 15

16. Autoscaler Case Study Target: find cmin to meet P99 latency goal Autoscaler Autoscaler Experimental Setup Production Setup cmin In Memory Cache In Memory Cache Web Server Web Server Database Database Downscaled Trace Original Trace Latency P99 < 500ms Latency P99 < 500ms 16

17. Autoscaler Case Study Target: find cmin to meet P99 latency goal Goal: 99th percentile latency (P99) < 500ms Policy cmin P99 latency (ms) AvgRate 0% 18147 T-span 40% 2757 Rand, TS-RRR 60% 557 RR, TS-LWL 70% 425 17

18. Conclusion Experimental Setup Production Setup In Memory Cache In Memory Cache Web Server Web Server User Traffic Database Database arrival time, url, payload size 150382450, www.foo.com, 500 193720475, www.bar.com, 20 215820849, www.abc.com, 472 222941037, www.efg.com, 82 230491837, www.gka.com, 129 249104962, www.ndw.com, 901 Downscale 11 ... - Downscaling: Lower a trace's load to run on a smaller setup - Key Idea: Use load balancing techniques for Downscaling - Real-world & synthetic trace experiments show TraceSplitter's superiority - Open-sourced code: github.com/smsajal/TraceSplitter 18