Bambiboom writes:
> I very much appreciate your very detailed reply and additional
> information regarding the performance of mainstream mechanical
Most welcome! Sorry there are still some data points missing, I'll
get them filled in when I can.
> drives relative to SSD's. The last several months I've become a
> Passmark Performance Test junkie, and perhaps there's a weighting
Although I've been collating typical benchmark numbers using HDTach,
AS-SSD, Atto, etc., I do try and perform various 'real world' tests,
such as the AE loading time. Benchmarks can be informative, but
remember the best benchmark of all is the specific application you
use, with a typical dataset. Did you notice the 'access time' test
I've been using on the disks page, a search involving thousands
of files?
http://www.sgidepot.co.uk/diskdata.html#ACCESS
It reveals some surprises, though it does also show the CPU
bottleneck of a 600MHz Fuel becoming more of a factor (SSDs run
even quicker with a Fuel/900). The data is of course mainly aimed
at SGI users, but it's nonetheless still useful in a general
sense (differences between models should be similar when tested on
a different system, unless the bottleneck lies elsewhere), and
very relevant to those with other SCSI systems, UNIX machines such
as SUN, IBM, HP, etc. What I might do is add a 2nd column to the
access time test results, where the same test is run on a quad-1GHz
Tezro instead, to show the effect of CPU performance on the results.
And I really need to add some other SSD models to that list...
> rating where the mechnical drives become SSD's. Having a fast disk
> system seems to be coincidental with higher CPU and memory scores.
It does help to have a good base system that can support the I/O
potential of an SSD; as always, a balanced system is best.
> The fastest systems seem to have what I think of as "direct
> injection" PCIe drives- I see a lot of RevoDrive 3 X2's in top
PCIe SSDs do offer very high sequential I/O rates, but you wouldn't
really notice any difference between a single premium SSD and a PCIe
SSD for general tasks, app loading, etc. By premium I mean the
Samsung 840 Pro, OCZ Vertex4/Vector, Corsair Neutron/GTX, Crucial
M500, etc. Actually, in many cases you probably wouldn't notice the
difference between a mainstream and premium SSD either, eg. the
Samsung 840 vs. 840 Pro, or Vertex3 vs. Vertex4.
It's the sequential I/O where PCIe SSDs can really shine, and
unlike mechanical RAID0 they won't barf if one does happen to hit
them with smaller request sizes. Do any of your tasks involve
manipulating datasets that consists of large single files? Editing
uncompressed video is a typical example. If you have to deal with
large datasets that also includes lots of smaller files, then a
PCIe SSD could really fly, though I'm sure others would point out
that you could just setup 2 normal SSDs using onboard RAID0, or
4 in RAID10, and achieve similar speeds in most cases.
However, I've not heard good things about the realibility of PCIe
SSDs. I certainly wouldn't use one as a system drive yet, but if
you do employ such a thing for general storage & RAID then some
kind of redundancy/backup is I would say essential, even if that's
just in the form of a single 2TB Enterprise SATA acting as a manual
backup.
> fast disk and real world applications' use.
This is the missing element in so many comments I read. Big numbers
are great, but if they don't lead to measurable target application
performance improvements, then who cares?
> slot, I will give the LSI SAS3080X a spin- literally. I didn't
Remember that card has no onboard cache. Do you have any external
storage at all? If so then the equivalent card is the 3800X. My
Fuel has a 3442X-R, so it has both internal & external links (backup
unit is external), but my Tezro has two 3800Xs since Tezro has no
extra internal drive bays.
> really expect the 3080X in the T5400 to meet the LSI chart- I'd be
> very happy if it performed at 50% those claimed rates. When I make
> transfers from partition to partition or from RE4 to Barracuda, Win
> Explr reports never more than 30MB/s !
Running tests on my T7500 might be a tad unfair as it's a different
chipset I think, ie. it'd produce overly optimistic numbers compared
to your T5400. It could be that your bottleneck is rather different.
Still, in theory, a SAS card with mechanical RAID0/1/10 should help
(don't use RAID5), combined with SSD(s) for system disk and/or apps.
I recommend RAID10 if you need RAID at all, ie. redundancy, but
with no sacrifice on performance. Best of both worlds. RAID5 means
too much in the way of CPU and other issues.
> I've toyed with the idea of RAID 0 and 1- and their combination-
> 10- but haven't taken the plunge. ...
I found it's easy to setup. I took screenshots of the tests I
ran, with file names that describe the test config. They're not
on my site, but if you like I can send them to you; find my
contact page by Googling "SGI Ian", or just PM me.
> brother's architectural office- new 2.93GHz Core 2 Duo to replace
> the 1.86, a Quadro FX 1700 (512MB) to replace a Quadro 550 (128MB),
> 64-Bit XP Pro, improve RAM from 2 to 4, set up a RAID 1. Total
> cost- $100.
Should give a nice speed bump. 8) Replace the system disk with
a used 120GB Vertex2E/3 off eBay, then he'll be really happy.
😀
> I see mention on your site of Irix and that brought back a flood of
Heh, I won't yabber too much about IRIX here as I'm sure it'll bore
other readers to death; feel free to PM/email if you'd like to chat
more about that. I have lots of SGIs including various high-end
systems (eg. 36-CPU Onyx3800), but my main desktop is a 900MHz Fuel.
> Adding 2MB RAM to have the maximum 4MB cost $180 -that's $9,000/GB
> and the replacement 540MB HD was $570- which at that rate would
> make a 1TB cost over $1,000,000.
😀
I remember seeing a quote in 1995 for a 256MB upgrade to a POWER
Challenge server, for 21500 UKP. And that was academic pricing
without VAT. But hey, the
Onyx I have in my garage was $1.5M when
new; three owners later (including BMW and Ford) I bought it for less
than the cost of a 3770K.
😀
> installed a number of Silicon Graphic Indys and I was jealous of
> their 100MHz processors and 16MB RAM, though they were more than
Hmm, low-spec Indys... SGI messed them up by bloating the kernel
rather a lot. Within a few years, 16MB was not enough to run IRIX
properly. IRIX 5.3 needs 32MB+, 6.2 needs 64MB+, 6.5 needs 128MB+.
I ran a student lab of 20+ Indys for a few years; insufficient RAM and
disk space was a real pain (the dept. got a 'deal' on a larger number
of Indys with a lower spec; big mistake, and typical of bad sales done
by resellers).
> on a drawing area and a popup viewport would display a detail block
> with annotations- astounding to me then. I thought I'd time-traveled
> to the future! I think that Indigo setup was in the $20,000+ range.
That's rather like what got me started, but I went in at the deep
end; the first SGI I used in 1993 was at Marconi Simulation, a quad-
CPU Onyx RE2 4RM4 rack with Sirius Video, running a tank simulator
on a 10' HD projection screen, ie. full 1920x1200 output, 48bit RGBA
with 8x subsample AA. The system had 256MB RAM, a huge amount back
then. At uni most systems were still 486 PCs; 3D on PCs didn't exist
at all yet. I was hooked. Here's some nostalgia for you:
http://www.sgidepot.co.uk/sgi.html#APPS
> A bit more about my system in case you have disk system or test
>
> Technical Designer, WdPft Office, MS Office. Tasks include 3D
> models, industrial design assemblies (up to 6,000 parts), renderings
> of same, Patent applications and drawings, architecture, and
Have you analysed what the system is doing when you're running
these various tasks? eg. monitor system usage re CPU, RAM, gfx,
disk, etc.? Process Explorer can show disk I/O, and of course there
are lots of benchmark programs for individual subsystem testing.
One thing to note, a single 6-core 3930K (possibly even without
being oc'd) would leave that old Dell in the dust. My Dell T7500
is nice, but it's not remotely a match for my 3930K. Beyond a
certain point, upgrading an old DDR2 Dell is probably not
worthwhile, as the bottlenecks become cheaper and/or easier to
solve by switching to a newer platform.
Certainly, switching to an SSD as a system drive would help your
workflow, I can guarantee that. I'd recommend an 840 Pro, Vector or
Vertex4. And as mentioned, you can always use a separate SSD to hold
installed program data if need be, to offset the disproportionately
higher costs of SSDs larger than 256GB. Mind you, given the T5400's
mbd limitations, I think you'd be delighted with the difference
you'd see even with just two Samsung 840 250GB SSDs for system disk
+ apps, or a couple of used Vertex3s like the ones I keep nabbing
when I can (bought several MAX IOPS units for good prices recently).
I struck lucky though last week, bagged two new
256GB Vectors on
eBay for 150 UKP each from different sellers. 8)
> descriptive industrial design proposal/descriptions. I believe the
> T5400 disk system (along with the slow memory) is a serious weak
> link in the chain.
Memory may indeed be a bottleneck. Having DDR2 rather than DDR is
at least something (trust me, a P4-based Dell 650 would be even
worse), but at some point it'll hold you back. Are you able to
adjust the RAM speed in the BIOS? If so, you could run tests with
the RAM at 667 vs. 400, etc. If there's a marked speed difference,
then one could infer that DDR2/800 would be better than 667, in
which case - assuming the T5400 doesn't support 800 speed - moving
to a newer DDR3 platform might be a wise move.
As for CPU speed, here's a simple example: run the Cinebench 11.529
CPU benchmark on your system, let me know what you get. My T7500
gives 10.90 (that's two 3.2GHz 4-core XEON X5570s with 24GB
DDR3/1333), while my 3930K at stock speed gives 11.13; oc'd, the
3930K gave 14.45 with the RAM running at DDR3/2133. Even a
comparatively lesser cost oc'd 4-core SB came very close to the
Dell, eg. 5GHz 2700K gives 9.86. See:
http://www.sgidepot.co.uk/misc/tests-jj.txt
> Again thanks for contributing to the mechanical HD / SSD debate.
Most welcome!
> intuition is that very fast disk systems along with GPU computing
Can any of your tasks exploit CUDA? I'll be testing CS6 shortly
with 4 x GTX 460 on various mbds, should be interesting.
Ian.
PS. I never had time to add results from the first 3930K AE system I
built to my results pages, but here's a CPU-Z link and a spec summary:
http://valid.canardpc.com/show_oc.php?id=2656382
http://www.sgidepot.co.uk/misc/aepc.txt
. And yes, difference between ssd and hdd gets considerably bigger when you don't wait for a bloated system to startup.
Facebook
Twitter
Instagram