[squid-users] cache performance: flash drive substitute vs. fast hard drive

2008-10-07 Thread Chuck Kollars
Anybody have performance experience (or benchmark results) putting Squid's 
cache on a Flash Drive? 

Devices that plug into a disk cable but that contain only what you'd find in a 
thumb drive are available. They have zero latency and they have much faster 
transfer speed than a moving disk. On the other hand they don't have any 
internal cache memory; even small repetetive accesses always go directly to the 
flash memory. (A regular hard drive typically has 4-32MB cache memory, so 
although overall access is only as fast as the disk spins, a few repetetive 
accesses can be very fast.) How do these two opposing tendencies (better 
average transfer rate but no internal cache memory) net out with Squid's cache 
access pattern?

For a Squid cache, am I better off buying a small but really fast hard drive, 
or one of these flash drive substitutes? 

-Chuck Kollars


  


Re: [squid-users] cache performance: flash drive substitute vs. fast hard drive

2008-10-07 Thread Joel Jaeggli
Chuck Kollars wrote:
 Anybody have performance experience (or benchmark results) putting
 Squid's cache on a Flash Drive?
 
 Devices that plug into a disk cable but that contain only what you'd
 find in a thumb drive are available. They have zero latency and they

They have no rotational latency however they're far from zero latency
devices... the fastest examples you can get now are in the ~80-100usec
range instead of the 8-15ms range.

 have much faster transfer speed than a moving disk. 

only in high end parts... many of the ones you see in laptops are
actually quite a bit slower than high-end winchester disks.

 On the other hand
 they don't have any internal cache memory;

Not a generalization that can be made, some enterprise models need
battery or capacitor backed write caches to order write erase cycles for
wear leveling. In general there's little point in having a read cache
however in places where it makes sense, some devices in fact do. .

 even small repetetive
 accesses always go directly to the flash memory. 

high repetive or extremely fragmented writes may be treated differently
by the controllers state machine eg by block shadowing so that large
regions don't have to be constantly rewritten for small writes.

 (A regular hard
 drive typically has 4-32MB cache memory, so although overall access
 is only as fast as the disk spins, a few repetetive accesses can be
 very fast.) How do these two opposing tendencies (better average
 transfer rate but no internal cache memory) net out with Squid's
 cache access pattern?

you're going to have the benchmark a particular variant in order to come
to grips with how that nets out... The 16GB sata ssd's I'm using from
last year in some security appliances are 1/2 the the speed reading and
1/4th of the speed writing as an analogous 10k rpm 2.5 sas disk in the
same box. Compared to a 4200rpm fujitsu ruggedized disk on the same
platform they are faster. Looking at the intel x25-m sata disk you can
see what a difference a year makes.

 For a Squid cache, am I better off buying a small but really fast
 hard drive, or one of these flash drive substitutes?

The other part of the equation is the ssd is still around an order of
magnitude or more per gigabyte more costly than the sas/sata winchester
drive, which is non-trivial when you're talking $700 or so for 80GB of
genuinely faster flash. If the alternative were buying 7x300GB 10k rpm
sas disks the flash route is a lot spendier for the equivalent capacity.

 -Chuck Kollars