Automated load balancing in the ATLAS high-performance storage software

Transcription

1

2 Automated load balancing in the ATLAS high-performance storage software Fabrice Le Go 1 Wainer Vandelli 1 On behalf of the ATLAS Collaboration 1 CERN May 25th, 2017

3 The ATLAS Experiment 3 / 20

4 ATLAS Trigger and Data Acquisition System 4 / 20

5 Data Logger System HBA1 Server 1 HBA2 Transient storage system to: Decouple online and oine operations Cope with disruption of permanent storage service or its connection Scale-out system, currently: 4 local-attached storage, 2 servers each 500 HDs, 430 TB, 8 GB/s Fully-redundant: no data loss caused in 2016 Server 2 HBA1 HBA2 Controller1 Controller2 Disks Expansion Disks Expansion Disks Expansion Disks Expansion HBA: Host Bus Adapter 5 / 20

6 Software Distributed in-house application (C++) Tasks: Receive selected events data Write data to disks Compute data checksum: le-by-le Adler32 Data-driven: events are distributed in classes called streams One le by stream New streams appear roughly every minute Stream distribution may vary rapidly Workload not uniform at all, cannot be fairly distributed Multi-threaded through task-oriented framework Typical Stream Bandwidth Distribution Another independent application sends the data to permanent storage 6 / 20

7 Threading Model Input threads IT1 IT2 ITN Events in streams Writing Manager creates File to thread bookkeeping S0 OT1 S1 OT2 S2 E0 E2 E7 E17 E3 E14 E10 E1... S0 S1 S0 S2 S0 S2 S0 S1 S1 S2 S2 S1 Writing Task event to 1 file Output threads File B OT1 assigned to OT2 File A File C OTM File A E0 E7 E3 E10... File B E0 E3 E2 E1... File C E2 E7 E17E The workload distribution is controlled by the le-to-thread assignment policy The application performance can be limited by one thread 7 / 20

8 Current Assignment Policy Round-robin: each new le is assigned to the next thread in a circular thread buer Simple implementation, very low overhead Deterministic behavior but events come with no specic order: non-deterministic assignment of les to thread The application's instantaneous performance is not predictable: The assignment of major streams to the same thread will degrade the application performance 8 / 20

9 Problem Modication in the operational conditions: higher throughput, dierent stream distribution Peak throughput 1.4 GB/s 3.2 GB/s S1: 80 % S1': 70 % Stream distribution S2: 6 % S2': 7 % S3: 3 % S3': 5 % Random assignment of major streams to the same thread will now degrade the application performance Synthetic test conrmed performance degradation: Conditions Writing rate Performance Loss No joint assignment 865 MB/s reference S1' and S2' together 797 MB/s - 8 % S1', S2' and S3' together 760 MB/s - 12 % 9 / 20

10 Weighted Assignment Policy A new workload distribution strategy was needed to restore performance and predictability Requirements: Data-driven: e.g. cannot assume any pattern in stream distribution Responsive: must cope with rapid evolution of stream distribution Low CPU and memory footprint Idea: Compute a load for thread: last-n-second sliding window of amount of processed data Assign a new le to the thread with the lowest load 10 / 20

11 Weighted Assignment Policy: Step 1 Threads load vs. time with assignments Zoom on the assignments: problematic in red Real-time load is ineective for close-enough assignments Reducing sliding window length: but cannot be too small, would be too sensitive to local uctuations (typical: 5 seconds) Another component needed to be added: Compute a load for the streams: same sliding-window amount of processed data by class of streams Add the stream load to the thread load upon assignment 11 / 20

12 Weighted Assignment Policy: Step 2 Threads load vs. time with assignments Zoom on the assignments Decisions are reected immediately: the likelihood of a thread to be selected again just after decision is inverse proportional to the load of the assigned stream 12 / 20

13 Testing Test in controlled environment with emulated data ow: Stream distribution and upstream event processing time emulated from 2016 monitoring data No wrong decision for + 40-hour runs Policy Writing rate Performance Gain Round-robin 865 MB/s reference Weighted 882 MB/s + 2 % Test on the actual ATLAS TDAQ infrastructure Used during ATLAS commissioning tests and cosmic data taking sessions 13 / 20

14 Conclusion The transient storage system of ATLAS TDAQ is a key component enabling for decoupling of online and oine operations Its workload is heavily unbalanced and cannot be fairly distributed In 2016 a new strategy was required to handle recent changes in operation conditions New workload distribution strategy: sensitive and self-adaptive to fast-evolving operation conditions and modications of the event selection process Validated in both test and production environments: proved to better use the parallel processing capabilities of modern CPUs for our workload This development will be part of the 2017 data-taking session 14 / 20

15

16 2015 Real-time Streams Writing Rate Figure: Instantaneous stream bandwidth for data collected on 28/10/2015. All streams are shown and each line represent a dierent stream. The highest line labeled "Global" is the sum of all streams representing the total bandwidth of selected events data. 16 / 20

17 2015 Stream Bandwidth Distribution Figure: Stream bandwidth distribution for data collected on 28/10/2015. Each bar represent the fraction of the total bandwidth for one stream over the considered period. 17 / 20

18 2015 Stream Bandwidth Distribution Figure: Bandwidth distribution between dierent streams for data collected on 28/10/2015. The four highest bandwidth streams are shown seperately and all other streams are summed together as "Other streams". 18 / 20

19 2016 Real-time Streams Writing Rate Figure: Instantaneous stream bandwidth for data collected on 24 and 25/10/2016. All streams are shown and each line represent a dierent stream. The highest line labeled "Global" is the sum of all streams representing the total bandwidth of selected events data. 19 / 20

20 2016 Stream Bandwidth Distribution Figure: Stream bandwidth distribution for data collected on 24 and 25/10/2016. Each bar represent the fraction of the total bandwidth for one stream over the considered period. 20 / 20