TEST: USB2 vs. firewire performance external harddisk

[Rob Lohman]

As some of you might know I never trusted my firewire cards to
give me the maximum claimed bandwidth of 400 mbit/s (or mbps).
This should translate to a rough maximum of 40 MB/s including
room for some overhead.

I was in the possession of a Maxtor 5000DV external firewire/
USB2 harddisk but it had recently died. Last week I received
a replacement from Maxtor and decided to run some tests with
the harddisk to see how it performed.

Disclaimer: I've only run read tests since I don't want to
destroy the data on it. I also don't claim to have used the
best testing program out there or the best cleanroom approach
to testing. Yes, most applications where closed on my two test
machines, but not all and both had active netwerk connections.
I only wanted to see what performance I was getting and compare
it to the different interfaces.

I tested with the following harddisk:

• Maxtor 5000DV v1.00.00
• Capacity: 160 GB
• Rotational speed: 7200 RPM
• Memory cache: 8 MB
• Supported interfaces: 400 mbit firewire & USB2

I used the following software/ OS etc. on both machines:

• Windows XP Professional Service Pack 1
• Internet Explorer 6.0 SP1
• AIDA32 Personal version 3.93 for system information and:
• Disk Benchmark plugin version 1.10 for benchmarking (comes with AIDA32)
• All partitations on all harddisks are formatted in NTFS

Edit on april 29th, 2004:
It looks like AIDA32 is no longer on that site. The guy moved his
product to a new company where you can download it as EVEREST.
Only problem is that it doesn't include the plugin to do the disk
benchmarking anymore.

Test system 1:

• Homebuilt tower
• AMD Athlon XP 2200+ (1800 MHz), 512 MB DDR PC3200
• ASUS A7V8X-X mainbord, VIA VT8377 Apollo KT400 chipset, Award bios (august 6th 2003)
• Firewire & USB2 ports
• nVidia GeForce 2 MX400 video, 64 MB, AGP 4x
• VIA AC'97 onboard sound

Test system 2:

• DELL Latitude C800 LAPTOP
• Intel Mobile Pentium IIIe 850 MHz, 256 MB SDRAM PC133
• Intel Solano i815 chipset, Phoenix bios (september 20th 2001)
• Only a firewire port
• onboard ATI Rage Mobility M4/128 video, 16 MB, AGP 4x
• onboard ESS ES1983S Maestro-3i sound

The laptop belongs to the company I work for and they have
ordered a new DELL Inspiron 9100 laptop for me which should
arrive within a few weeks. I will update this report with the
test results of that machine.

I ran the following tests with the Disk Benchmark plugin:

1. Quick Linear Read

   This test is designed to measure the linear (sequential) reading performance in a shorter duration than the Linear Read test. This test does not read all the data from the surface of the device, but to achieve a 1/10 duration in compared to the Linear Read test it reads only the 1/10 of the surface of the device

2. Average Access

   This test is designed to measure the data access performance of the storage device by reading small data blocks from random locations on the surface of the device

3. Buffered Read

   This test is designed to measure the performance of the storage device interface (when it's possible) by reading only the very beginning of the surface repeatedly. This test will work only with storage devices equipped with built-in cache memory of at least 64 KB.

As indicated before I did not run any write tests nor did I
run the "Random Read Test" since it produced very similar
results nor did I run the "Full Linear Read" test since the
sampling of the quick test was good enough.

[Rob Lohman]

On to the results I hear you scream. Okay. The first table
lists only the averages so you can get a quick indication:

Averages, linear read test:

• System 1 [homebuilt tower] - USB2: 31.4 MB/s (16% cpu usage)
• System 1 [homebuilt tower] - firewire: 33.7 MB/s (4% cpu usage)
• System 2 [DELL laptop] - firewire: 32.6 MB/s (6% cpu usage)

Averages, average access test:

• System 1 [homebuilt tower] - USB2: 20.5 ms (3% cpu usage)
• System 1 [homebuilt tower] - firewire: 20.5 ms (2% cpu usage)
• System 2 [DELL laptop] - firewire: 19.8 ms (2% cpu usage)

Averages, buffered read test:

• System 1 [homebuilt tower] - USB2: 32.1 MB/s (16% cpu usage)
• System 1 [homebuilt tower] - firewire: 34.1 MB/s (4% cpu usage)
• System 2 [DELL laptop] - firewire: 33.0 MB/s (5% cpu usage)

Already we can see some interesting results from this test.

Firewire seems to be the winner. It has higher throughput
on every test versus USB2 (on my test system!) and also less
CPU usage! Interesting. I would've expected the reverse to be
true for at least CPU usage. That was a feeling I was having
which just has been proven wrong.

We can also conclude that I'm using the full bandwidth of my
USB2 and firewire busses since there is a marginal increase
in the buffered read tests.

The last conclusion is that the busses do loose some from
the theoretical maximum of 40 MB/s (firewire) and 48 MB/s
(USB2). But I do think 30+ MB/s for reading isn't too shabby
taking overhead into account. Perhaps a superiour interface
card could increase it a bit, but I doubt it will be that
much. So this already debunks my thoughts on my slow firewire
drive (unless there was something more substantially wrong
with my previous drive, ofcourse)

[Rob Lohman]

Now onto the full results of the tests for some interesting
other conclusion:

LINEAR READ test:

System 1 [homebuilt tower] - USB2

• Minimum: 20.4 MB/s - 3% cpu usage
• Maximum: 32.8 MB/s - 33% cpu usage
• Average: 31.4 MB/s - 16% cpu usage

System 1 [homebuilt tower] - firewire

• Minimum: 29.3 MB/s - 3% cpu usage
• Maximum: 34.5 MB/s - 10% cpu usage
• Average: 33.7 MB/s - 4% cpu usage

System 2 [DELL laptop] - firewire

• Minimum: 29.7 MB/s - 1% cpu usage
• Maximum: 33.0 MB/s - 15% cpu usage
• Average: 32.6 MB/s - 6% cpu usage

AVERAGE ACCESS test:

System 1 [homebuilt tower] - USB2

• Minimum: 18.8 ms - 1% cpu usage
• Maximum: 22.9 ms - 9% cpu usage
• Average: 20.5 ms - 3% cpu usage

System 1 [homebuilt tower] - firewire

• Minimum: 18.7 ms - 1% cpu usage
• Maximum: 22.2 ms - 5% cpu usage
• Average: 20.5 ms - 2% cpu usage

System 2 [DELL laptop] - firewire

• Minimum: 18.2 ms - 0% cpu usage
• Maximum: 22.0 ms - 5% cpu usage
• Average: 19.8 ms - 2% cpu usage

BUFFERED READ test:

System 1 [homebuilt tower] - USB2

• Minimum: 7.7 MB/s - 1% cpu usage
• Maximum: 32.8 MB/s - 35% cpu usage
• Average: 32.1 MB/s - 16% cpu usage

System 1 [homebuilt tower] - firewire

• Minimum: 32.2 MB/s - 1% cpu usage
• Maximum: 34.4 MB/s - 12% cpu usage
• Average: 34.1 MB/s - 4% cpu usage

System 2 [DELL laptop] - firewire

• Minimum: 30.8 MB/s - 1% cpu usage
• Maximum: 33.3 MB/s - 23% cpu usage
• Average: 33.0 MB/s - 5% cpu usage

There are a couple of extreme numbers in relation to the
other numbers, like: 33% cpu usage (this was a small
fluke probably due to something doing something in the
background) and 7.7 MB/s (again I'd say something else
was doing something because it was only one small little
drop in the overal graph).

There are two very interesting conclusions to make from
these more detailed sets of data besides the conclusions
I already made from the average numbers above. And that
is the following:

First, it seems like the drive is actually (a lot) faster
than the firewire or USB2 bus can support. Howso? Because
we really don't see any dip in performance throughput.
Normally when you test a drive that is as fast or slower
than the interface you will gradually see the throughput
falloff when the heads travel over the platters. This is
due to the drive running at the same rotational speed on
the outer and inner tracks of the drive will the amount
of data changes (since the drive is smaller at the inside
track than at the outer edges).

We don't see such a fallof in our numbers which means
that the slowest throughput of this drive is still faster
than the interface can read. Interesting! This is where
firewire-800 (with a 800 mbit or roughly 80 MB/s
intercace) might get real interesting.

The second observation that I made was that the seektime
(average access time) on the drive was lower on my
laptop (which is better!) than on my other faster machine
(with USB2 and firewire basically turning over the same
numbers). Which is a bit weird since normally seek times
don't have anything to do with the computer to which the
drive is attached. I'm at a loss for explaining this.

Summed up:

So the definite overal conclusion seems to be that built-
in USB2 and firewire ports seem to do the claimed speeds
(more or less). I haven't had a change to run the same
tests on a cheap USB2 and firewire card and I might do
so in the near future when I'll be re-running the test
on my new laptop as well.

The second big conclusion is that firewire seems to have
an edge over USB2 (definitely over USB1 ofcourse!) in
terms of transfer rate, although a small one, and CPU usage.

And the last conclusion is that the current drives seem
to out perform the current available external interfaces
(not counting SCSI, ofcourse).

So there you have it!