[gpfsug-discuss] Rewriting existing files is incredibly slow
Jonathan Buzzard
jonathan.buzzard at strath.ac.uk
Mon Mar 11 17:35:12 GMT 2024
On 11/03/2024 16:43, Peter Hruška wrote:
> Hello,
>
> We've encountered yet another performance flaw. We have a GPFS
> filesystem mounted using GPFS binaries on Windows. When we try to
> rewrite a file on the GPFS filesystem rewriting speed is much slower
> than writing to a new file. The difference ratio we measured is about
> 3.5 times. From the task manager it is visible that there is excessive
> amount of reading from the network when rewriting. This is even visible
> on the NDS server as io activity. However when rewriting on a linux
> client there are no reads while rewriting. Has anyone encountered such
> problems?
> To replicate the issue is is possidle to run fio twice with the same
> configuration to achieve rewriting or to run iozone. Both tools return
> similar outputs.
Kind of yes.
What are you using to "rewrite" the file?
What we saw initially over Samba was certain Microsoft applications when
rewriting a file had truly abysmal performance. The same application
when saving the same document to a new file and the performance was as
expected.
After much digging into it the cause (it was weeks of person effort) it
was determined to be down to the application writing the file *one* byte
at a time. Basically some idiot C++ developer at Microsoft decided to
ignore the C++ library because it has "bugs" and write their own
formatted output routines.
It was not noticeable saving to a local disk, but the instant you tried
saving to a network drive the performance was truly awful. Basically the
increased latency of single character IO was the issue.
Note that the issue was not confined to Samba and GPFS, as we verified
the same abysmal performance with a Windows 2008 R2 server running on
File and Print Sharing on NTFS on bare metal hardware. It was also not
confined to just Windows the same awful performance happened on Macs
too. In fact that is where it first came to light.
Might be the cause of your problem, might not.
JAB.
--
Jonathan A. Buzzard Tel: +44141-5483420
HPC System Administrator, ARCHIE-WeSt.
University of Strathclyde, John Anderson Building, Glasgow. G4 0NG
More information about the gpfsug-discuss
mailing list