Re: Tracing Filesystem Accesses
On 5/15/2011 4:21 AM, Boyd Stephen Smith Jr. wrote:
> In <4DCEC70C.email@example.com>, Stan Hoeppner wrote:
>> On 5/14/2011 11:02 AM, Rainer Dorsch wrote:
>>> My Main reason for not doing it was that I do run the SSD partitionless
>>> to avoid partition alignment issues.
>> There are no partition alignment issues with SSDs. SSDs have no
>> cylinders, no heads, no mechanical parts. Native sector size is 512B,
>> access time to every sector on the device is uniform--non issue.
> SSDs do benefit slightly from having partitions aligned with pages or blocks.
Do you have some test results, or a link, showing the performance
difference between aligned and non-aligned current generation SSDs, to
back this claim?
> Those pages are generally somewhere between 4KiB and 64KiB in size. Blocks
> are larger, some getting up to 8MiB. <http://www.anandtech.com/show/2738/5>
> and <http://www.nuclex.org/blog/personal/80-aligning-an-ssd-on-linux> seem to
> be decent references.
Neither of those articles demonstrate an SSD performance deficit due to
the read-modify-write misalignment penalty, which is what hammers mech
disk performance. SSDs don't have to wait for the platter to come back
around for the write operation after the read, there is no head required
to seek across the platter, and they don't write the data back to the
same physical location to boot, but to a clean cell block. Thus the
write is taking place in a few hundred microseconds, not multiple
> Traditionally, being unable to TRIM has caused reduced performance over time
> on SSDs. However, SSDs have such a performance increase over SMDs that I
> doubt misalignment would be noticed very much.
Totally agree. Even if all cells are 'dirty' due to a full drive, and
require an erase before a write, the erase takes only 1-2ms, and
multiple erases can occur in parallel. Something most everyone tends to
forget is that fragmentation on a mech drive will drop your performance
by a factor of 10-100 or more depending on application profile and/or
multiuser system load. And yes, all Linux filesystems DO fragment,
contrary to popular myth, very badly with mbox files (IceDove), or any
long lived constantly appended or modified files such as databases.
Filesystem fragmentation simply doesn't exist on solid state storage.
Given currently available information, I feel the performance drop due
to age issue with SSDs is over blown a bit.