No description, website, or topics provided.
Switch branches/tags
Nothing to show
Clone or download
Permalink
Failed to load latest commit information.
LICENSE Initial commit Aug 18, 2015
README.md Recommend noop scheduler system-wide Aug 31, 2017
testoverlay.py Add script to test for overlayfs Aug 24, 2015

README.md

ranisalt's guide to a faster computer

First of all, take your time to measure the bottlenecks. You can check if your boot time is reasonable with the systemd-analyze tool:

$ systemd-analyze
Startup finished in 3.104s (firmware) + 50ms (loader) + 914ms (kernel) + 1.409s (userspace) = 5.479s

And you can go deeper and check what takes most time to load from userspace passing the parameter blame:

$ systemd-analyze blame
           879ms psd-resync.service
           601ms docker.service
           319ms psd.service
           285ms dev-sdb3.device
           277ms wicd.service
           276ms user@1000.service
           135ms dev-sda2.swap
            86ms systemd-journald.service
(...)

General recommendations

  • Disable unnecessary services. You can check what are the major hogs with systemd-analyze blame, and it will list processes taking boot time, sorted by heavier first. Then, mask them with systemctl mask and you're ready, and systemctl unmask if something goes wrong. You will look forward to disabling Bluetooth, infrared, and even firewalls if you are not concerned with it.

  • Use a lighter display manager. SLiM is always a good option, LightDM is good too. You can also go with no DM at all, shaving off precious time. If your distro comes with, consider disabling Plymouth too (we're going FAST, not EYE-CANDY).

  • Use a lighter bootloader. GRUB can take a big time just running scripts. If you have only one boot option or you can use the UEFI bootloader, you can use lighter options such as Syslinux and Gummiboot, or if you are savvy you can give a shot to EFI stub.

  • Use a lighter desktop environment (or none). Run away from GNOME and KDE if you are pursuing performance, as those giants can consume up to one gigabyte of memory and lots of CPU cycles. Opt to a lighter alternative, such as XFCE and LXDE, or even better, give up completely on DEs and use only a window manager, such as dwm, i3 (my personal choice) or xmonad.

  • Avoid using fancy disk settings, such as LVM, if you don't need to. Also, don't use encrypted partitions (/home is fine). Also remember to set your SATA disks to run in AHCI mode, not IDE.

  • Replace your HDD with SSD. This is probably the simplest option. You don't need to run your entire system on SSD to get a boost, you only need to have your boot files (/boot partition) and executables (/usr), although personally I only have /home and swap on HDD.

  • Use lighter software always. From personal experience, I found that wicd is faster and lighter on resources than NetworkManager. Firefox tends to use less memory than Chrome. mpv is a simple but amazingly complete media player, usually better than Totem and VLC. Consider learning vim instead of an IDE or graphical editors. Also, you do not need an office suite since Google Docs is great.

If you followed the guide religiously up to here, you are ready to dig into more command line stuff. Don't be scared, I won't brick your computer.

Finer system tuning

  • Decrease output during boot. Your system might spit out a ton of information during boot, but really it's too much and too fast to care. Go to your bootloader configuration (Google it) and append the following to the kernel command line:

    quiet loglevel=3 splash
    

    It removes most of the messages and sets the log level to display only error messages. You can read the boot output later using dmesg and journalctl.


  • Use faster disk schedulers. If you have a SSD, you can change the scheduler since the default one, CFQ, accounts for seek time of HDDs, which do not exist on SSD. I recommend the deadline scheduler, but the noop scheduler can be faster, especially on boot. To configure it, first create the file /etc/udev.d/60-scheduler.rules and insert:

    # set deadline scheduler for non-rotating disks
    ACTION=="add|change", KERNEL=="sdb", ATTR{queue/rotational}=="0", ATTR{queue/scheduler}="<scheduler>"
    

    Replacing sdb with your disk (you can get its name with lsblk) and <scheduler> with your favorite scheduler (e.g. my configuration. You can also replace the disk scheduler on boot, in which case there will be less overhead on boot times, adding the following to your kernel command line (like in the previous tip):

    elevator=<scheduler>
    

    This post has some benchmarks with cfq, deadline and noop and concludes native SATA queueing (called NCQ) probably does a slightly better job than kernel schedulers, so noop might have the best throughput. Replacing accordingly.


  • Finer tuning of /etc/fstab. First of all, you can use the noatime option on ext filesystems to prevent updating access time of files, thus not writing to the disk every time you touch a file (although touch still works). You can also specify discard on SSD partitions to activate TRIM (Google it), and commit=60 on SSD to sync to disk less frequently. For example, the full line for my / partition:

    /dev/sdb3	/	ext4	discard,noatime,commit=60	0	0
    

    Addendum: this blog post claims that discard might have a negative impact. As always, your mileage may vary.


  • Optimize your initrd. On Arch Linux, you can modify /etc/mkinitcpio.conf and it is very well documented, and there's a pretty explained guide here (don't mess with it if you use LUKS). You can change the COMPRESSION to cat on SSD to generate a larger initrd in favor of faster (instant) decompression, a fine trade-off around 10 times faster than gzip, or you can use lz4 compression which can compress more than 30% and has almost instant decompression (benchmark). On other distros, refer to the official guides and manuals.

  • Use profile-sync-daemon. It is a great tool by @graysky2 to manage your browser profile on tmpfs, and it both increases performance and reduces disk wear. If supported, use it in "overlayfs" mode to minimize sync delay. To check if your system supports it, run the supplied script testoverlay.py, or run the following manually:

    $ zgrep OVERLAY /proc/config.gz

    On Fedora (kernels that do not export /proc/config.gz), you need to run the following:

    $ grep OVERLAY /boot/config-$(uname -r)

    And it should output either CONFIG_OVERLAY_FS=m or CONFIG_OVERLAY_FS=y, else you need to recompile your kernel. Then, edit /etc/psd.conf and uncomment USE_OVERLAYFS="yes". Reboot and check.


  • Replace bash by dash on boot. This is a tricky one. dash is a very slim alternative to bash and it can be used on boot to shave off some milliseconds. You need to redirect /usr/bin/sh to /usr/bin/dash if it was linked to bash. Check it first:

    $ ls -og /usr/bin/sh
    lrwxrwxrwx 1 4 Aug 14 03:03 /usr/bin/sh -> bash

    So yes, it was linking bash. Go to /usr/bin and change it:

    $ cd /usr/bin
    $ ln -s dash sh

    And you are good to go. Just remeber to update your scripts if they rely heavily on bash and are configured to use /usr/bin/sh.


  • Avoid swapping (or disable at all). Swap is a major bottleneck. You will not want to have it on SSD since it will wear out your SSD, and use a too big space if you have a small SSD like me, but HDD is very slow and when your system starts thrashing, it'll feel like hell.

    Create a file under /etc/sysctl.d named 99-swappiness.conf and add the following:

    # Do less swapping
    vm.dirty_ratio = 6
    vm.dirty_background_ratio = 3
    vm.dirty_writeback_centisecs = 1500
    

    Those configurations are responsible for controlling how often the kernel swaps (documentation), I use them on a 6 GB system with a 8 GB swap partition and this is a good value. Here's a good article on how swappiness works, I recommend the reading.

There's more to come.

License

This work is under public domain.