IORateLimiter's auto tune strategy is not optimal
Nobody has claimed this yet.
- Dominant language
- C++
- Stars
- 1k
- Forks
- 423
- Avg merge
- 1d 15h
- Merged PRs (30d)
- 24
Description
Enhancement
Reproduce steps
- load chbenchmark 1500 into a cluster with tiflash
- set the io rate limiter with config as follow
[storage]
[storage.io_rate_limit]
background_read_weight = 0
background_write_weight = 20
foreground_read_weight = 0
foreground_write_weight = 80
max_bytes_per_sec = 209715200
- run the chbenchmark tp with 100 concurrency
tiup bench ch --host 10.2.12.81 -P 8080 --warehouses 1500 run -D chbenchmark -T 100 -t 0 --time 30m
foreground write and background write
We can see that bg_write alloc bytes is much more than fg_write alloc bytes. The current auto tune logic tends to tune more write_bytes limit for bg jobs rather than fg jobs. Thus causing fg write pending. This is not as expected.
When the total bytes per seconds almost reach the limit, we should first ensure the requirements for fg jobs rather than bg jobs.
Read alloc mismatch with read request
As shown in the below figure 1 and figure 2, whether it is foreground reading or background reading, "read alloc" is always much smaller than "read req". This may be caused by the file system's PageCache.
However, as shown in figure 3, it still cause unexpected rate limit on the fg read or bg read, though the disk throughput is still free.
Sub-taks:
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the IORateLimiter auto-tune logic and reproduce the behavior using the provided TiFlash configuration and chbenchmark command. Compare foreground and background read/write allocation with request rates and disk throughput; done means foreground work is prioritized near the total limit and available disk capacity is not unexpectedly throttled.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- databases, performance
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100