Three Simple Tweaks for Better SSD Performance

Dmitri Popov

Productivity Sauce

Nov 25, 2009 GMT
Dmitri Popov

As I explained in the previous post, replacing my notebook's hard disk with an SSD significantly improved the overall system performance -- even without any additional tweaking. But there are also a couple of simple tricks that can boost performance even further. The first one is to disable the sreadahead service. The sreadahead tool helps to speed up the boot process with conventional hard disks, but it actually slows the boot with SSDs. To disable the service, open the sreadahead.conf file for editing using your preferred text editor:

sudo nano /etc/init/sreadahead.conf

Comment then the following line:

exec /sbin/sreadahead -t 0

Next trick is to add the elevator=noop kernel boot parameter to disable the elevator scheduler. On Ubuntu 9.10, open the grub.cfg file for editing:

sudo nano /boot/grub/grub.cfg

Add then the elevator=noop parameter as follows:

linux     /boot/vmlinuz-2.6.31-15-generic root=UUID=b5c7bed7-58f1-4d03-88f4-15db4e367fa0 ro   quiet splash elevator=noop

This scheduler is used to read and write data from the hard disk sequentially. Since an SSD is not a conventional hard disk, disabling the elevator scheduler significantly improves the read and write performance of your SSD.

Finally, you might want to set the file system mount option to noatime. To do this, edit the /etc/fstab file, so it looks something like this:

/dev/sda1    /  ext4   noatime,errors=remount-ro   0       1

Adding the noatime option eliminates the need for the system to make writes to the file system for files which are simply being read -- or in other words, this means faster file access and less disk wear.

That's it. Now reboot your machine, and you should notice faster boot and better performance.

Comments

  • noatime is safe for most application

    ctime, mtime, atime dates back to the foundations of Unix is available in most POSIX-compliant operating systems.

    atime (access time) is updated whenever a file system object is read -- even from cache.

    ctime (ionode modification time) gets updated whenever the file system object's ionode is changed (e.g., file/directory is created, ownership is changed, permissions are changed, file/directory is deleted).

    mtime (object modification time) gets updated whenever the file system object is modified (e.g., truncated, appended, overwritten).

    Well-written backup applications use the mtime to flag modified files for progressive, incremental, and differential backup modes. If the backup application also preserves ownership & permissions, it should also look at ctime. atime is no use whatsoever here -- backups are interested about file changes -- not reads.

    atime has a few edge cases. Some systems audit file reads to track usage. tmpwatch removes aged files from the /tmp directory based on atime. mail-notify looks at atime to notify users about new mail. Cache & proxy servers once used atime for pruning objects from their cache, but this has been largely replaced by tree indices that can weigh multiple factors (e.g., number of requests vs. aging).

    For the most part noatime is completely irrelevant. If in doubt, read the opinions of the holy trinity of Linux development -- Linus Torvalds, Andrew Morton, and Ingo Molnar:
    http://kerneltrap.org/node/14148
  • noatime may break some applications like backup tools

    If you are using Karmic, relatime will be safer bet. Note that this option might not work well with older kernels.
  • noatime implies nodiratime

    A common confusion - http://lwn.net/Articles/245002/ .
  • Anon

    Disabling readahead may not be a win (at boot) as it will fetch blocks earlier than they would have been fetched otherwise and (depending on the readahead implementation) only when disk I/O would otherwise be idle. Thus more is done I/O is done in parallel earlier in the boot process so programs wait even less leading to a faster boot reagardless of the fact the underlying disk does not have a read seek penalty...

    Something similar may apply to general readahead too - if you are going to need those blocks later sequential readahead can still be a win as you will simply wait less later (at a point when you would have had to stall).
  • sreadahead

    note that karmic doesn't use sreadahead, it uses ureadahead which automatically detects ssds and performs appropriately.
  • journaling on SSD is also a bad idea

    By doubling the number of writes in certain operations, it increases the wear on the SSD, and slows things down even more. Ext2 may be old (and well-debugged), but it is still a strong contender for certain workloads. Until btrfs reaches production quality, ext2 is probably the best way to go for SSD.
  • fstab

    One more tip, add "nodiratime":

    /dev/sda1 / ext4 noatime,nodiratime,errors=remount-ro 0 1
  • setting readahead values for just one drive

    If you want to set the readahead for just one device you can adjust it using the blockdev command like so;

    blockdev --setra 0 /dev/sda

    Obviously you'll want to put this command in a init.d file or similar...

    p.s. this is only for 2.6 kernels
  • Test results

    It would be very interesting to see speed results. how big performance increase exactly we will get form this tune up?
  • Re: GRUB configuration changes

    @Juergen Pabel: gus3 has already mentioned this technique, but thanks a lot anyway!
  • Re: Quality information

    @Andrea R: Actually, I wrote both posts on the same day, so the fact that you mentioned the tweaks covered in the article is just a coincidence. Or maybe not, since all the tweaks are more or less well-known.

    There are many ways to skin a cat, and I don't pretend that my way is the best one. However, I did try all the modifications myself and they worked.

    You are right, though. I should have thanked you for commenting on my post. I'm always grateful to people who find time to share their ideas and knowledge with others. Thank you!
  • GRUB configuration changes

    Hi,

    on Ubuntu 9.10 I recommend editing /etc/default/grub instead of /boot/grub/grub.cfg. Any changes made to /boot/grub/grub.cfg directly will be lost the next time update-grub2 is executed (like when a new kernel or kernel-patch is installed via software update). So, by changing the line

    GRUB_CMDLINE_LINUX_DEFAULT=quiet splash"

    to

    GRUB_CMDLINE_LINUX_DEFAULT=quiet splash elevator=0"

    in /etc/default/grub and then running update-grub2 one can make this change more persistent.

    Juergen Pabel
  • Re: editing grub.cfg is not advised in Ubuntu

    @gus3: You are absolutely right. Thanks for the tip!
  • thanks

    Thanks for the hints, and thanks to the people who commented on them. I notice an improvement.
  • Quality information

    Hilarious, you basically took my comment to a previous entry and made it into a whole new post. You forgot the filesystem suggestion and put the noatime trick instead, which is not ssd-specific.

    You also do things in a terribly dirty way.
    Learn to disable a service. Uh uninstall readahead may be even esier. That and editing grub configuration properly.

    Why don't you give it some time, get some real experience and test the modifications thoroughly then start filling the web with quality information?
    Also, next time at least say "thank you".
  • editing grub.cfg is not advised in Ubuntu

    Any edits to /boot/grub/grub.cfg will probably be lost at the next kernel upgrade.

    The preferred way to change boot parameters in Ubuntu 9.10 is by editing /etc/default/grub, altering the line that starts with GRUB_CMDLINE_LINUX_DEFAULT=... to suit your needs. Once you have saved it, run update-grub to apply the changes to /boot/grub/grub.cfg. Verify that the changes are in place, and then reboot.

    Or, in the specific case in this article, just "echo noop > /sys/block/DEVICE/queue/scheduler" to change the elevator without rebooting, substituting the correct device name for DEVICE.
comments powered by Disqus

Issue 170/2015

Buy this issue as a PDF

Digital Issue: Price $9.99
(incl. VAT)

News