New article about software RAID and PM

44 views
Skip to first unread message

Eduardo Berrocal

unread,
Oct 26, 2020, 11:02:03 PM10/26/20
to pmem

Anton Gavriliuk

unread,
Oct 27, 2020, 4:01:24 AM10/27/20
to Eduardo Berrocal, pmem
Hi all

Please keep in mind, that in the section

Appendix C: Volume Configuration

They using fsdax (not sector) mode.  So Db/App working on such setup must be aware of 8kb atomicity.

I'm also going to check what is currently faster mdadm or lvm raid for write intensive workloads.

Anton

Anton Gavriliuk

unread,
Oct 27, 2020, 4:11:18 AM10/27/20
to Eduardo Berrocal, pmem
8kb -> 8 byte 

вт, 27 окт. 2020 г. в 10:01, Anton Gavriliuk <antos...@gmail.com>:

Eduardo Berrocal

unread,
Oct 27, 2020, 12:54:56 PM10/27/20
to pmem
Thanks Anton for pointing that out. I also have a paragraph in the article:

"All workloads in this experiment had DAX enabled. I/O against DAX means that applications that rely on write atomicity at the block/sector level can get their data corrupted by torn blocks/sectors in the event of a crash or power failure (x86 architecture only guarantees to write atomically 8 bytes at a time during memory copies). The lack of block write atomicity doesn’t affect the file system’s metadata (the file system takes care of that). For a more extended discussion about this topic, and to see some performance numbers, please read my previous article Speeding Up I/O Workloads with Intel® Optane™ Persistent Memory Modules."

Reply all
Reply to author
Forward
0 new messages