xfs
[Top] [All Lists]

XFS - configuration for multi-thread high-speed streams

To: "xfs@xxxxxxxxxxx" <xfs@xxxxxxxxxxx>
Subject: XFS - configuration for multi-thread high-speed streams
From: Mike Jensen <Mike.Jensen@xxxxxxxxxxx>
Date: Tue, 15 Oct 2013 14:45:28 -0600
Accept-language: en-US
Acceptlanguage: en-US
Delivered-to: xfs@xxxxxxxxxxx
Thread-index: Ac7J5DhWlIY0Qa5LSrekxIwvrmjfHA==
Thread-topic: XFS - configuration for multi-thread high-speed streams

Hi Group - 1st time post, looks like just use the email address I have seen from other posts

If the setup/info/questions below are not suitable for this list, is there another xfs list available for these??

 

Thanks

mike

 

Configuration –

 

-          RH 6.4 on HP G8 server w/16GB mem (could add a lot more)

-          12Gb SAS connections (2) to storage

-          Storage

-          96 LFF HDD’s (4TB 7K NLSAS) presently organized as 8 x 10+2 R6 raidsets

-          8 volumes mapped to host ports (MPIO engaged)

-          Mdadm used to stripe (/dev/dm*) LUNs into a single md device (approx. 300+TB file system space)

-          Mkfs.xfs used to lay down file system

-           

-          Workload

-          Application creates file of specified length in dir of mount point, IO’s then issued to file(s), each file receives 1 stream of IO

-          1 x high speed stream 1000MB/s Seq W (IO’s arrive as 512KB -=> 4MB frame would be 8 x 512KB IO’s)

-          2 x med-speed streams 200MB/s Seq W “

-          30-50 low speed streams 10MB/s Seq W “

 

Objectives/Questions –

 

-          Would like to optimize xfs/mount parameters to make maximal use of storage assets

-          Thinking of using 15K HDD R1 set for log files – would this get all/most of metadata or just a subset?  Right now seeing metadata writes arriving with data writes and want to peel them off the 7K HDD’s

-          Would using sub-directories for each file ensure that each subdir/file would be in it’s own alloc group? And that would keep the streaming data in a separate AG – could help the storage more efficiently destage the writes from cache to disk

 

<Prev in Thread] Current Thread [Next in Thread>