>> I logged this bug:
>http://bugs.opensolaris.org/view_bug.do?bug_id=6720764
>
>Ages ago.
>
>I just did the dd test with an 8 MB block size and obtained about 16 MB/sec,
>so this (apparently) confirms the problem is related to pcfs, though in the
>past I also had bad performance using UFS....
>
Specifically in USB or specifically on flash media?
There are many different types of readers and types of media.
E.g., I used to have a VIA EPIA (now broken) and its flash reader
worked at 12MB/s (it's on the motherboard, but still connected with
USB 2.0)
I have 3 more CF card readers:
white box: 5MB/s
Targus: 8MB/s
New Lexar: 17MB/s (and 27MB/s on a new UDMA card)
(The other cards used are Sandisk Extreme III (older edition, no UDMA);
see also http://www.robgalbraith.com/bins/camera_multi_page.asp?cid=6007-9392
UFS on USB connected spinning media seems reasonable, but nowhere near
a direct connection (SATA/IDE).
Flash media is different; and I'm not sure whether UFS does very well;
the amount of actual IOPS is important for flashmedia and UFS doesn't
know how to optimize.
I'm not sure the kernel ordered IOPs send to flash media; the only thing
it should do is coalesce IOPs which would really help.
When you have USB key, it's difficult to say how quick it would be (dd
is a fine tool, here). Cheap and/or older USB keys can be much worse,
especially when there are a lot of IOPs going on.
Clearly there's a lot of things we need to do: particularly bringing
the number of IOPs down needed for a specific operation.
Try reading from the same USB key using a 2K blocksize and compare
with the 16MB/s you get with 1MB blocksize.
Casper
