Thursday, September 1, 2016

Everything you Need to Know about Hard Drive Vibration

Hard Disk Drives (HDD's) are one of the most impressive and important electromechanical devices ever created. When I think about it, it is really amazing that these things actually work, let alone work so well!  These disks must quickly and precisely position heads slightly above very narrow tracks on rapidly spinning platters. If there is an error in the angular position of the head, it will not be positioned above the correct sector and therefore cannot correctly read or write data. 

One thing that can interfere with this amazingly fine and fast positioning is vibration. Vibration can cause small variations in the position of the head with respect to the track, so it can be an enemy of the proper functioning of HDD's.  But the mechanics of HDD vibration is an obscure subject, and as a result, there is an aura of mystery surrounding vibration. As part of my job, I've spent a lot of time and effort to understand vibration and to engineer our systems to deal with it.  I'm hoping to share some of that with you to help demystify the subject.

HDD Structure and Scale

The figure below shows the structure of a typical HDD, showing platters, tracks, heads, sectors, and cylinders (sets of corresponding tracks on platter surfaces).

To get a sense of how small all of this really is, let’s do some back of a napkin calculations.
  • A WD Purple 8TB HDD has approximately 248 million tracks cut up into just north of 15 billion sectors across all platters.
  • Let’s assume there are 8 platters in there and all are double-sided. That means there are ~15.5 million tracks on each surface of a platter.
  • Assuming 3.5” outer diameter and 1.75” inside diameter, the actual usable space of the platter has a radius of 0.875”.
  • That leaves us with 17.7 Million tracks per inch.
  • Considering the thickness of a piece of standard letter paper is 0.004” that means you could fit 70,800 tracks on the edge of a piece of paper.
That really is some tiny dimensions we're working with here.

Once you see how tiny the positional tolerances are, it becomes obvious that even the slightest vibration could move the arm (with respect to the platter) while reading or writing, resulting in an IO miss. As a result of this miss, the HDD must let the disk spin one revolution and try again. This greatly increases the time to complete the read/write, under extreme vibration, the operation may be delayed for several disk revolutions. 

These read/write delays are the root of all vibration-induced performance degradation.

Vibration Sources

I'm going to ignore vibration from sources outside the computer since it's more an issue for laptops and desktops, I work with servers which are mounted in racks and generally not subject to external vibration.

That leaves us with two main two main sources of internal vibrations: fans, and the HDD's themselves.

Fan Vibration - Fans are a source of vibration but good quality balanced fans have less vibration and a large heavy solidly built server, such as our 45 Drives' Storinator, will reduce fan vibration. If the manufacturer pays attention, fan vibration can be well under the threshold that HDD's can handle.  

Now, let's talk about vibration caused by the HDD's themselves. It comes two sources:  platter wobble, and reaction to head seek acceleration.

Platter wobble - Caused by a slight imbalance in the rotating platters. It is a readily identifiable sinusoidal vibration at a frequency that exactly corresponds to the rotational speed of the platter (90 Hz for 5400 RPM, 120 Hz for 7200 RPM, etc.). The dominant imbalance is in-plane so vibration is in the X-Y plane (see diagram below).

Reaction to head seeks acceleration HDD heads are mounted on arms, which rotate to seek to new tracks. These seeks can happen very rapidly, upwards of 100 seeks/second, so the arm must accelerate quickly.  According to the laws of physics, a rapid acceleration requires a large force, and each action has an equal and opposite reaction. Thus a rotational force is applied to the HDD body. Although the HDD body is massive relative to the head and arm, the reaction force is sufficient to slightly move the HDD, causing vibration in the X-Y plane.

A Direct Look at HDD Vibration  

The theoretical explanations are just fine, but it is much more interesting to actually see it. So we bought an accelerometer chip, wired its 3 axis outputs through op-amps. These were fed into a storage scope so we could record waveforms. Then I attached the acceleration to a Western Digital Ae 6TB HDD which I picked because I happened to have it on hand, and happens to have typical vibration characteristics.   I then recorded HDD vibration under two conditions: free floating (actually suspended on elastics); and attached to a massive block of granite. These show the maximum and minimum magnitudes of internal HDD vibration (free-floating shows worst case, HDD mounted to granite shows the best case, whereas HDD's mounted in a real server are an intermediate case).

Free Floating HDD - Maximum Vibration

The next two graphs were captured while the HDD was free floating. This represented the worse case drive mounting scenario. The Y Axis is the acceleration in G's (1G =9.8m/s^2), and the x-axis is time.

The first graph the vibrations are measured along the X-Axis. The Blue trace is when the disk is powered on but sitting idle. You can see that there is a consistent sine wave with a frequency of 90Hz, This is platter wobble. Converting that to RPMs and you get 5400 RPM, exactly what the disk is rated to spin at. The green trace was taken while the disk is under a random write test. You can clearly see peak accelerations that result from the head seeking the next sector. 

Soft Limit - The yellow lines are a soft limit around ~0.22 Grms where performance degradation will start to become annoying (exceeding %10). The vibration level causes position error that is great enough to cause an IO miss but not big enough to cause a full on timeout between the disk and the disk controller.

Hard Limit - The red lines are a  hard limit at ~0.67Grms, where the HDD will experience non-recoverable errors and the HDD will not operate in a useful fashion at all.

The next graph is the same setup, but displaying Y-Axis data (recorded at a separate time).  The blue trace is under idle conditions and the green trace is a during a random write.  Again, yellow is the soft limit, red the hard limit.

I haven't shown the Z-axis because vibration is minimal.

HDD's Rigidly Mounted to Granite 

The next graph was captured while the HDD was mounted to a large granite slab, this represented a best case drive mounting scenario. The blue trace is the vibration in the X-Axis and the Green is the vibration in the Y-Axis. Yellow is the soft limit.  (I did not bother to include traces during seek as the pattern is exactly the same as in the free floating scenario.)  Note that scale is different.  The dramatic reduction in vibration is obvious.

Observations and Conclusions

Note that in the free floating case, vibration is mostly under the soft limit, but on head seek the Y-axis vibration peak is just at the soft limit. The limit is an RMS number, so therefore we can conclude that if all hard drives had to worry about was their own excitation, vibration-induced degradation is not a problem.

If we rigidly couple the HDD to anything massive, vibration is dramatically reduced. Therefore a single HDD mounted to a chassis with no other sources of vibration will perform well,without any vibration problems.

Unfortunately, in the real world, a server typically has multiple disks, a large PSU with fans, and multiple cases and CPU fans. These all create a vibration which superimposes.  Thus one must carefully design the chassis such that HDD's are coupled to it with sufficient rigidity; and that the chassis itself has sufficient mass and rigidity in order to suppress vibration.  All this must be measured carefully. I'll be writing another blog entry showing what this all looks like.

Free feel to leave a comment below or contact us at


No comments:

Post a Comment