Date: Fri, 18 Aug 2000 08:47:06 -0700 (PDT) From: Matt Dillon <dillon@earth.backplane.com> To: "Allen Pulsifer" <pulsifer@mediaone.net> Cc: <freebsd-hackers@FreeBSD.ORG> Subject: Re: RE: what to do about /tmp Message-ID: <200008181547.IAA55501@earth.backplane.com> References: <NBBBJNDFEKPEHPFCLNLHMEIDHNAA.pulsifer@mediaone.net>
next in thread | previous in thread | raw e-mail | index | archive | help
:
:Matt,
:
:I noticed /tmp and /var/tmp are used slightly differently, for example, the
:contents of /tmp are not expected to survive a reboot, while /var/tmp is
:used for files you might need after a crash, like vi recovery files.
This may have been true at one time in the past, but these days
/tmp and /var/tmp tend to be used interchangeably. You can't make
those sorts of assumptions.
:One advantage of mounting /tmp with MFS is that you don't have use up another
:disk partition. Performance-wise, I would think data moves something
:like this:
:
:/tmp on MFS:
: application data --> kernel VM --> paged to disk
:
:/tmp on FFS:
: application data --> kernel file buffer --> flushed to disk
:
:Is that basically correct? What makes MFS less efficient than FFS?
:What would be the implications of creating a 512 MB swap partition
:and mounting an MFS /tmp?
:
:Thanks,
:
:Allen
There are several problems with MFS:
* The active data set is cached in memory TWICE rather then just
once. That is, for each page of file data two pages is taken up
of memory.
* While it is true that MFS will swap unused data, performance under
heavy load will go to pot because MFS does not in any way attempt
to sequence the data to swap.
MFS may appear fast under light loads, but performance degrades
drastically (to much worse then what you get with FFS) under
heavy memory loads.
* MFS eats a 'random' amount of swap, up to the size of the MFS
disk. Under medium loads where paging is taking place, the
swap backing store for MFS becomes fragmented and performance
goes to pot.
There are two solutions:
* First, you can use vnconfig to create a swap-backed filesystem
and you can reserve the necessary swap wspace contiguously,
resulting in performance which will be as good or better then
a normal FFS filesystem (because your swap space may be striped
across several disks).
* Second, you can use a normal FFS partition.
In either case enabling softupdates on the vnconfig or normal
FFS partition should result in good overall peformance without
the load penalty.
With both solutions it is a good idea to make /var/tmp the partition
and make /tmp a softlink to /var/tmp.
Using MD instead of MFS is typically not a great choice. While it is
true that MD only caches data once, the data is not swap-backed so
what data MD does allocate is allocated permanently. This is useful
in certain system configurations but not useful in the general case.
-Matt
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200008181547.IAA55501>