[SERVER-64510] Evaluate YCSB as unloaded workload Created: 15/Mar/22  Updated: 08/Apr/22  Resolved: 08/Apr/22

Status: Closed
Project: Core Server
Component/s: None
Affects Version/s: None
Fix Version/s: None

Type: Task Priority: Major - P3
Reporter: Daniel Gomez Ferro Assignee: Jordi Olivares Provencio
Resolution: Done Votes: 0
Labels: None
Remaining Estimate: Not Specified
Time Spent: Not Specified
Original Estimate: Not Specified

Attachments: PNG File 100-updates-t2.png     PNG File 50-50-t2.png     PNG File Screenshot 2022-03-31 at 16.08.34.png     PNG File ycsb_0_100.png     PNG File ycsb_100_0.png     PNG File ycsb_50_50.png     PNG File ycsb_95_5.png    
Sprint: Execution Team 2022-04-04, Execution Team 2022-04-18
Participants:

 Description   

We want to choose some workloads that aren't overloading the system to verify the overhead of the FIFO queue doesn't impact uncontended cases.

YCSB could be such a workload, we need to make sure the current configuration isn't overloading the system by looking at:

  • CPU utilization
  • Eviction
  • Read/write tickets
  • Flow control


 Comments   
Comment by Jordi Olivares Provencio [ 01/Apr/22 ]

Ah, I see, I've taken a look at it and I'm in agreement with daniel.gomezferro on this. The workload seems CPU bound as there's no maximum concurrency reached in Wiredtiger. My only concern would be in the version with 50/50 and 100% updates that do eviction, but they both seem to have the exact same profile and even values.

50/50

100% Updates

Comment by Daniel Gomez Ferro [ 31/Mar/22 ]

jordi.olivares-provencio initially I was looking at the "small scale" YCSB for this case, called industry_benchmarks, sorry for not being more clear about that.

Comment by Jordi Olivares Provencio [ 31/Mar/22 ]

I took a look and ycsb_60gb seems to vary wildly in performance characteristics at least according to this patch build in a 3 node replica set. In particular the only configuration that somewhat offers an "unloaded" view is the 95% reads + 5% updates. But even that seems to contain a non-trivial amount of cache evictions. As to the rest of configurations:

  • The 100% reads view sufers from read tickets contention
  • The 50/50 as well as 100% updates modes hit Flow Control
Comment by Daniel Gomez Ferro [ 31/Mar/22 ]

My initial assessment is that the workload is CPU bound but there's no other resource exhaustion.

Generated at Thu Feb 08 06:00:31 UTC 2024 using Jira 9.7.1#970001-sha1:2222b88b221c4928ef0de3161136cc90c8356a66.