Tweaking Solid State Drive Performance On Linux

perlow writes "While Solid State Drives are expensive and shouldn't be used exclusively for primary storage, they perform exceedingly well for things like MySQL databases, provided you tweak your kernel, BIOS, and filesystems accordingly. Here's a few tips to get excellent performance out of your new $500-$900 investment on a Linux system."

Comments

EeePC

[lukrop]

I'm interested how useful this could be for my EeePC since it only uses SSDs. let's see

SSD

[jrwr00]

SSD Drives are nice and all, but a Major MYSQL DB would eat a flash disk alive, all those writes, even if the drive supports 10 billion writes, for a big mysql DB, it could eat that in a year, You will still need a shit ton of ram cache for the main DB so all the minor writes dont kill the disk, there is the fact of data loss in the power loss... all writes stored in memory will be lost... HDs are slow yes, but SSD is just not there just yet

Converted my garage PC to a SSD-alike

[karnal]

I've just purchased a 4gb transcend 300x card (udma5) and an addonics IDE to CF adapter. This article comes along at a great time, since scouring the net for information on "flash linux" usually results in some article regarding USB installations, which are about 2x (or more, depending on drive used) slower than this solution.

The only thing that this PC runs is minor browsing when looking up a car repair, or Amarok (with a MySQL backend.) Regarding the mysql backend, it doesn't seem like there will be enough action on the disk to even begin to worry about the flash failing.

In addition, the transcend I purchased (and most high end == costly) is an SLC device, which I've read is faster as well as can sustain more writes over the lifetime of the drive. I see this as worthwhile in a system such as this, as I want it to be an "install and forget" type of system. Using a spinning platter in the variable humidity and temperature conditions just doesn't seem to make sense - especially when most of the media used would be in a server in my basement.

I'm also looking into network booting; however the PC that I've slated for use in the garage would probably require a bootable floppy/cd since it's old enough to not be bootable from an installed NIC. Maybe I just need to add a newer NIC - but I've not spent a lot of time researching that.

If anyone has other articles I could enhance my knowledge with - and possibly doing something similar to my setup - I'd love to read them.

Wait, what?

[Briareos]

He recommends disabling journalling and using RAID instead?

So exactly how will a RAID make sure the filesystem metadata is still intact when I yank out the power cable for fun and no profit, as opposed to using a filesystem with a journal?

Sheesh... that's just begging for an accident to happen.

np: Yello - You Gotta Say Yes To Another Excess (Orb Goes The Weasel Mix) (Auntie Aubrey's Excursions Beyond The Call Of Duty (Disc 2))

Re:Wait, what?

[Florian Weimer]

He also recommends to turn on write-back caching. I'm not really sure if this is save. The cache on the SSD is probably not battery-backed.

Re:Wait, what?

[schon]

He also says that the 'noatime' mount option tells the kernel not to read the atime... noatime tells the kernel not to write atimes to the fs.

sweet jebus, is there *anything* technically correct in that article?

Partitioning

[mickwd]

I'm surprised I've heard very little about using Unix/Linux partitioning to get the best out of SSDs.

Seems to me that the best use of an SSD on a normal system is to buy a smallish one (say 16GB) and use it for the read-mainly partitions: say /usr, /opt, maybe /lib.

It would be good to get users' "dot" files in there too. Maybe create a /homedot on the SSD and symlink /home/myname/.example to /homedot/myname/.example.

Even if this doesn't make your applications run much faster, the faster read and seek times are going to make the machine boot faster, load applications faster (especially including the desktop environments, if user directories like .kde and .gnome are on SSD) and compile code faster (with /usr/include, etc on SSD).

Re:Partitioning

[imroy]

...partitioning is a good idea for any disk. It prevents fragmentation...

How do you figure that? Modern Unix filesystems actively avoid fragmentation by being careful about allocating blocks to files. With more free space, a filesystem has more options as to where to place a new file. It's better to have one big filesystem with 10G free, than four or five with only about 2-3G free each.

Re:Partitioning

[Siffy]

Fragmentation is of little concern with SSDs. The time needed to read out of order bits on a rotating media is much higher as we know, but reading bits out of order in memory or on SSD is almost identical to reading the bit beside it due to the low seek times. Also, partitioning schemes that try to keep the most often accessed files at the beginning of the media lose their edge for the same reason. A SSD has the same performance characteristics from the beginning to the end of the drive from 1% to 99% used capacity. Partitioning could actually shorten the lifespan of the drive, because it could prevent the wear-leveling algorithms from doing their job properly. You'd likely have the most often written to partition fail much sooner than the rest of the drive.

SSD & MySQL benchmarks

[damg]

Here some SSD benchmarks for MySQL with a conclusion of "with the relatively low cost of the technology, you could net 10X+ performance increase on your database servers for under $2000."

Don't assume things

[jgoemat]

Do the math. Most flash media are good for at least 100,000 writes. They also use wearing algorithms so that each block averages out to about the same. Even if you try to write to a certain block over and over, the algorithms take over and move blocks that are not written very often to those locations, so blocks end up being written to about the same.

With that in mind, let's take the 64 gb model for $899 as an example. Let's say you have a huge workload and are writing at the max of 35MB/sec. At 100,000 writes, 64GB gives you 6.4e15 bytes that can be written before the disk wears out. At 3.5e7 bytes per second, that comes to 1.83e8 seconds. There are 3.16e7 seconds in a year.

That means that at THE MAX write rate and only 100,000 write cycles (not 10 billion), the drive should last at least 5.8 years. Unless you are simply writing to a logging database, you won't be using anywhere near the max write speed and the drive should last decades.