On Wed, Dec 17, 2014 at 10:26 PM, Dan Egli <[email protected]> wrote:
> On December 16, 2014, [email protected] wrote:
>
>> i did an experiment a couple years ago where i booted a
>
>> windows machine off an iscsi target connected over gigabit ethernet.
>
>> loading games and regular desktop usage was almost as good as if the
>
>> hard drive was connected to the local system.
>
>
>
> That's actually impressive. I didn't think Gigabit went that fast. I recall
> reading somewhere that a base rule of thumb when speaking of speeds over
> the network was to take the advertised speed and quarter it to get the
> actual usage speed. 25MB/sec isn't very fast at all these days. Even
> mechanical SSDs frequently read at 150MB/sec and write at 100MB/sec if the
> benchmarks I've seen are any indication. I may have to do some tests.
I would bet that the "feel" of running a system over a network drive
vs. a local drive depends a lot more on latency than throughput, and a
lightly loaded gigabit network has pretty nice latency
characteristics. On my current network at work, an ICMP round-trip is
about .2ms, which is only 10x slower than via the loopback interface,
which is about .02ms. Spinning disks have *minimum* seek times in the
same order of magnitude as network RTT, and the disk's onboard cache
will be cold at boot. A network drive is likely to be a SSD, a RAID
array, or at least have a hot cache. So certain access patterns are
probably *faster* via iSCSI on a lightly loaded network than via a
local spinning disk, or at least there's a potential for being faster.
There are plenty of other variables involved, so YMMV.
--Levi
/*
PLUG: http://plug.org, #utah on irc.freenode.net
Unsubscribe: http://plug.org/mailman/options/plug
Don't fear the penguin.
*/
