[MDEV-6664] Windows IOPS issue (slow read) Created: 2014-08-29  Updated: 2014-09-04  Resolved: 2014-09-04

Status: Closed
Project: MariaDB Server
Component/s: None
Affects Version/s: 10.0.13
Fix Version/s: N/A

Type: Bug Priority: Major
Reporter: Simone Giordano Assignee: Unassigned
Resolution: Not a Bug Votes: 0
Labels: performance, windows
Environment:

Windows 7 x64, SSD disk


Attachments: PNG File Memory limited.PNG     PNG File PerfMon Disk MariaDB.PNG     PNG File PerfMon Disk SQL Server.PNG     PNG File Result MariaDB.PNG     PNG File Result SQL Server.PNG     Text File Table.sql    

 Description   

There is something strange on disk usage (IOPS) on Windows version.
I'm doing some tests porting an ERP archive from SQL Server to MariaDB.

The test archive is a ledger table with 22.800.000 rows, each rows has 69 columns for a total row length of 832 bytes.
There is a primary key which is an autoincrement.

On the same machine (i5 4 core, Samsung SSD 840) there is a big difference between SQL Server and MariaDB.
My test is with a query without indexes (ex. search for a specific value or GROUP BY and SUM) so the engine have to do a full table scan.
I need this type of test (obviously I can speed up the query with an index).
Each engine is limited to 2 GB of RAM (for MariaDB with innodb_buffer_pool_size).

Same query repeated a lot of times, takes about 30sec for a full table scan on SQL Server and about 90sec with MariaDB. 3x times!

Watching filesystem activity with Performance Monitor, I've seen that SQL Server can reach 400/450 MB/sec reading its MDF file.
MariaDB is stuck at 150 MB/sec to read IDB file.

I've tried with innodb_io_capacity parameter but without lucky.

In attachment you find table structure and some screenshots.



 Comments   
Comment by Sergei Golubchik [ 2014-09-04 ]

It doesn't look like a bug so far. InnoDB internal file layout is different, may be it's not well as good optimized for full table scans, but works much better for point lookups or ranges. Either way, I don't think it's a bug.

Generated at Thu Feb 08 07:13:40 UTC 2024 using Jira 8.20.16#820016-sha1:9d11dbea5f4be3d4cc21f03a88dd11d8c8687422.