HTPC build for Plex three part series:

1. Hardware build and Storage Spaces
2. Plex server and clients
3. Performance

I’ve had a single disk Home Theater PC (HTPC) running Plex for a while now but with my ReadyNas nearing the end of its useful life, I decided to kill two birds with one stone: build an updated HTPC with drive redundancy to also take over the functions of my aging NAS. My HTPC stays on 24/7 anyway so this makes perfect sense. My design goals for this were to build a small, quiet, low power, data redundant and sufficiently performing box that I can stuff into my TV cabinet. The extent of the duties for this HTPC will be video streaming and file sharing/ backup, no gaming. This series details my build, Plex server, Plex clients, Windows Storage Spaces, as well as the performance of the storage subsystem.

Performance

In these performance tests I want to focus on the main activities my HTPC will be undertaking: file read/ write operations and streaming. I tested using both Parity and Mirror Spaces to demonstrate the capabilities of each.

WD Red Baseline (Control)

First to set a baseline for WD Red 3TB HDD, here is the observed write performance with a single 40GB file write operation (from C:\ to Z:\). The copy took around 4.5 minutes and the drive topped out around 145 IOPS. Subsequent read performance tests top out around the same level. Considering these disks have been reported to spin at 5400 RPMs, frankly that’s really not too bad.

Windows reported ~142MB/s transfer speed:

These drives produce absolutely ZERO audible noise. No clicking, no head seeking noise, no rotational noise…nothing. With the cover off of this box, all that is audible is the 3 spinning fans (case, CPU, PSU). Under the TV cabinet, my build is dead silent.

File copy between Mirror and Parity Space

Test constants:

C:\ = SSD
N:\ = Mirror space
M:\ = Parity space

This test demonstrates the performance capabilities of a Parity Space using the WD Red disks. Disk performance copying multiple files from a mirror space (N:\) to a parity space (M:\) was fairly poor, averaging 33 write IOPS on the parity space. Observed read performance on the source 2-way mirror space was low at 25 IOPS on average, but that’s all it was being asked to do. This is somewhat expected however, as one of the caveats with parity resiliency is reduced write performance.

Once again, the Windows file copy dialog tracks very closely to the observed maximum write performance in Perfmon.

Power consumption of the system under test varied between 50 to 75 watts. Very good!!

Write performance of Mirror and Parity Spaces:

Test constants:

C:\ = SSD
D:\ = Mirror space
M:\ = Parity space

Here you can clearly see the write performance disparity between a two-way mirror and parity space. Copying a 4.5GB file from SSD (with very high read performance) to my two-way mirror space yielded very reasonable write performance according to the Windows file copy dialog. The overhead of the mirror operation appears to be costing around 20 IOPS.

The write activity of the physical disks participating in the mirror.

That same 4.5GB file written from SSD to my parity space (M:\) yielded significantly lower results (almost 3.5x lower):

Here is a more detailed view of the write performance against the parity space:

Read performance of Mirror and Parity Spaces:

Test constants:

C:\ = SSD
D:\ = Mirror space
M:\ = Parity space

Copying the same 4.5GB file from each space back to the SSD yielded better results overall. The read performance of the parity space during this operation is better but still below what the mirror space or even native drive are capable of:

Copying the file from the mirror space to the SSD was higher performing than the parity space and on par with native disk read performance, but the gap isn’t as wide here.

Plex Streaming Performance: Single File

Test constants:

-Physical disk 1 and 2 comprise a mirror Space
-Plex clients in use: Roku
-Media in use: single file, highly compatible, 1080p

Streaming being a highly sequential mostly read operation, there isn’t a great tax put upon the system resources. Multiple streams and transcoding complicate this a bit. The following chart shows the performance of a single video stream from my HTPC using a mirror Space to my Roku client. Interesting to note that the C:\ drive is busier than both disks in the mirror space, but all reads are sourced from a single disk.

Plex Streaming Performance: Single File, skipping around

Test constants:

-Physical disk 1, 2 and 3 comprise a mirror Space
-Plex clients in use: Chromecast
-Media in use: single file, highly compatible, 720p

Trying to force spaces to utilize all disks within a vDisk, I manually skipped around four times within a sample media file. Here you can see that with every color change within the performance chart is where I jumped to a different time within the media file. Almost every time I did this, Spaces offered up a different disk to be read from. The one exception is during the period between 7:40 and 7:43 where I performed a jump but the same disk was used. Otherwise, all three disks in  my mirror Space were used to read the file.

Plex Streaming Performance: Multiple Streams from One File

Test constants:

-Physical disk 1 and 2 comprise a mirror Space
-Plex clients in use: Roku, Chromecast, PS3 (DLNA)
-Media in use: single file, highly compatible, 1080p

To satisfy my curiosity, regardless of how infrequently this situation might actually occur, I wanted to see what the impacts were if I kicked off multiple streams of the the same file. I chose a mostly compatible file so we get the benefits of Direct Play here. Three simultaneous streams requiring transcoding would probably not work so well. There is a CPU spike when the third stream is launched and minor spikes on disk 2 for each stream event. You’ll note here that all reads are coming from disk 2. Looking at each TV screen individually and subjectively, the playback was completely normal.

Plex Streaming Performance: Multiple Streams from One File, 3 Disk Mirror

Test constants:

-Physical disk 1, 2 and 3 comprise a mirror Space
-Plex clients in use: Roku, Chromecast, PS3 (DLNA)
-Media in use: single file, highly compatible, 1080p

Just to see if there is any impact to performance, I ran this same test again but with three physical disks in my mirror Space. I would have expected my reads to be coming from at least two disks but as evident here, disk 1 is doing almost all of the work. Interesting to note that overall spikes are much lower now with read performance hovering around 1 IOPS. The takeaway here is that this third disk in my two-way mirror Space doesn’t add much from a performance perspective.

Plex Streaming Performance: Multiple Streams from Multiple Files

Test constants:

The following chart shows the effects when additional streams are added using three different media files, one requiring transcoding. You can see that system resource usage is relatively low with the first two streams, things change dramatically with the introduction of the third file. Transcoding means heavy CPU usage as evident here. Disk activity with 3 different files being accessed simultaneously jumps up quite a bit too. 

Zooming in here looking at the disk IO, you can see once again that almost all read requests are satisfied from a single disk. Hmm, where’s my striping performance??!

Here is the Plex transcoder in action:

Network Performance

Network performance of the Plex server with a single HD media stream is very reasonable. There are spikes north of 11Mbps but for most of the duration the  transmitting bandwidth stayed around 2Mbps.

Conclusions:

• The WD Red HDDs are a very quiet, very low power option that yield decent disk performance. Performance is more than acceptable for an HTPC that doubles as a NAS.
• Two-way mirror Spaces are the highest performing option in Storage Spaces but come with the highest disk cost as well (2x). Read performance is on par to that of a native disk with Spaces adding no discernable overhead, writes cost ~20 IOPS in overhead. If you have HDD capacity to spare, this is the resiliency option you want!
  • While writes clearly hit all disks in a mirror Space, observed read operations from streams are satisfied from a single disk in the Spaces pool. Jumping around within a media file does force reads from all disks in a Space.
  • 3 disks in a Storage Space really doesn’t buy you much other than additional capacity. With a 1-column two-way mirror you still only have two copies of data, on two physical disks. Read and write performance is not increased here.
• Parity spaces will work fine for Plex and video streaming in general, as video streaming is sequential and read performance is acceptable. Parity spaces suffer massively in reduced write performance so copying files to or moving files around on your HTPC will be much slower. The trade off for poorer write performance in a parity space is more overall usable capacity (1.5x consumed in a parity space vs 2x in a mirror space).
• Plex makes excellent use of system resources and can handle just about anything you can throw at it. The Plex server doesn’t even break a sweat with highly compatible media types, even with multiple streams. Transcoding is CPU intensive and should work fine for a single stream, multiple streams may tax your system and degrade the viewing experience.

This post originated at Exit | the | Fast | Lane – weestro.blogspot.com

Contrary to popular attempts to convince you otherwise – the PC is not dead. Fewer novice-level users may actually require a PC these days thanks to smart phones and tablets but I’m not quite ready to concede to “app store” style games and pecking out emails with my index fingers pressed on smudgy glass as my only means. I have smart phones, a tablet, several laptops, and yet I still find myself wanting and using a PC more than the others combined. Multi-displays serving rich graphics rendered from a discrete GPU, Chrome with 50 tabs opened at once, a few RDP sessions, streaming audio, RARs unpacked and copied in seconds, Visio, PowerPoint, Premier Pro, the latest DX11 game titles…you get the picture. Some things just work better with a mouse and keyboard and honestly, unless Minority Report ever becomes a reality (via Kinnect perhaps?) this will probably always be the case. My current PC is still humming along just fine with a 4-year old Windows 7 build on what was then, top of the line hardware. Ivy Bridge is here now, parts are more efficient, use less power, and are faster than ever. Time to upgrade.

Parts Selection

Now that I decided mATX was the way to go, it was time to assemble the Bill of Materials. Case selection was not an easy feat and I usually find myself just ahead of the curve, wanting what’s just about to come out. My last few ATX builds have been using Lian-Li cases which are high quality, aluminum, artful designs that do tend to cost more. Silverstone is another manufacturer of like ilk creating some of the highest reviewed products in this space. Taiwan seems to have the high-end PC case market cornered between these two companies.

I looked at a lot of cases over the course of several days seeking the smallest most functional mATX case I could find. I based my decision between 2 cases: The PC-V355B by Lian-Li and the Sugo SG09 by SilverStone. Lian-Li’s case is the closest thing out there to the Node 304 in looks, trouble is there are fitment limitations and design problems. The biggest restriction is the CPU cooler height at 100mm which rules out any high performance tower style cooler. Some people have reported that the motherboard once installed, does not sit flush to the IO shield in the back.  Another problem is that the PSU perches over the motherboard without even a shelf to support it! TweakTown’s review of this case highlighted that using a larger ATX PSU actually caused the rear of the case to flex. Not good. Shame because she’s pretty.

SilverStone on the other hand, created a case that is not only smaller but has no restrictions on CPU coolers or GPU lengths plus supports ATX sized PSUs in a proper shelf. The SG09 has received high praise from just about every review I read. Space efficient, quality construction, free flowing and small. Exactly what I’m looking for. The largest point of contention with the SG09 are its looks. At first glance I didn’t like it either. That front panel is actually plastic made to look like brushed aluminum. I can honestly say that I’ve stared at this thing so long now that it has really grown on me. SilverStone announced at CES a premature successor to the SG09 called the SG10. It’s essentially the same case with a flatly-styled aluminum front panel. Initially I was holding out for the SG10 which is rumored to ship some time in April 2013. The more I looked at the 2 cases the more I liked the SG09 and less so the SG10. I’m also not afraid to modify this thing if it ever bothers me. Functionally this case is very close to mATX perfection which at the end of the day is what matters most.

I didn’t find building in the SG09 terribly difficult, but I’ve been doing this for a while. Careful part selection is of utmost importance and having patience working in tight spaces is key. There is a big warning right on the box to dissuade novice builders.

For my power supply, considering this SFF build in a tight space, I chose a 140mm modular PSU from Silverstone: ST55F. Gold efficiency, clean power, fully modular, with a max of 600 watts, this PSU should provide everything I need and more. I also purchased the SilverStone PP05 short cable kit which, in hindsight, is still a good idea but I should have gone with a fewer higher quality PP07 cables instead. I used maybe 2 or 3 cables from the PP05 kit leaving the others to waste. Space is at a premium so the less you have to tuck inside of tight spaces the better! The review site JonnyGuru.com does an incredibly detailed job reviewing power supplies so definitely stop there first to check out any PSU contenders.

CPU, RAM and Mobo

I’ve long since abandoned AMD as the economy play in the CPU market for Intel who continues to have very little competition in the high-end CPU space. Ivy Bridge is no exception and is the smallest (22nm) most powerful hyperthreaded quad-core yet, with power consumption under 80 watts. More and more functionality is going on-chip with this “Tick” in Intel’s Tick-Tock cycle, getting a die shrink and enhanced graphics capabilities. I debated my CPU choice considering even the 3570 as clock cycles really don’t matter today as much as they once did. Overclockability, cores, cache, power consumption and instructions are all notable considerations but my decision was ultimately made up for me by price. 4 years ago I got an amazing deal on my 1st gen Core-i7 920 CPU from MicroCenter locally, this time around MicroCenter again did not disappoint. They were selling the high end 3770K CPU (K=unlocked) for $100 less than NewEgg, who are really the last bastion of online PC parts. The lesser chips that I was considering were more expensive than this high performer. This was really a no brainer.

The 3770K provides a ton of performance right out of the box. Pushing this chip further will be an easy albeit potentially unnecessary task.

CPU cooling is an area that has a lot of very good data to help make your selection, if opting to buy into the after-market. The stock Intel coolers are marginally ok, barely. If any overclocking is to be attempted at all, an after-market cooler should be a serious consideration. Noctua is one of my favorite brands of cooler with somewhat expensive but highly effective products including some of the best fans around. The NH-D14 is a MASSIVE cooler that is pretty much top dog right now. It occupies a tremendous amount of real estate so low profile RAM is a must in SFF. It’s also $80. For a fraction of that cost ($35), the Cooler Master Hyper 212 Evo delivers impressive results stacking well against its pricier counterparts and was my choice for this build. It is a somewhat rawer product with a direct contact style sleeve where the CPU rests directly against the flattened copper pipes of the cooler. Some minor prep work was required on the 212 Evo before installation, namely filling the cracks between the pipes with Arctic Silver 5.

I also replaced the cheapy Cooler Master fan with a high quality Noctua for silent operation and high performance. Here it is installed in the SG09. There is plenty of clearance for DIMMs with taller heat spreaders.

For motherboards, I’ve bought into the Asus high end for a very long time and have never been disappointed. Performance, value, stability, and longevity are all traits I’ve come to expect from Asus. ASRock has ramped up as a worthy competitor along with Gigabyte with many good reviews and satisfied customers, but at the $200 price point I felt no need to stray from Asus who has only left me satisfied in the past. The Maximus V Gene is the Asus answer for the high-end mATX z77 chipset market. Overclocking, customization, features, performance… everything I could want is in this board and for less than many of its high-end competitors.

Memory is another area of agony with so many choices on the market. One thing was clear at the outset: I wanted a 32GB 1600Mhz DDR3 kit (4 x 8GB). When building SFF the size of the heat sink on the RAM is a key consideration especially if you plan to run an after market CPU cooler. I wanted a kit with tight timings (CAS 8 or CAS 9), low voltage (1.35v) and a low profile to sit unobstructed in the DIMM channels. The Crucial Ballistix Tactical and Sport VLP products fit this bill perfectly. The Tactical LP DIMMs (CAS8) in yellow would have been my first choice but there was a lack of third-party supply for this build and buying directly from Crucial for much more really negated the value for these modules. I ended up going with the 32GB VLP kit (CAS9) in black and have absolutely no regrets. All head-to-head reviews peg these kits very close in performance with a slight edge going to the Tactical LP for overclocking. The Tactical LPs also have a slightly taller heat spreader and carry a higher price premium. The VLP (Very Low Profile) sit almost flush to the DIMM channels providing zero obstruction for any CPU cooler. These modules are seriously low profile and with Micron chips under the hood, you can absolutely trust their performance and longevity.

Both run very tight timings at the stock 1.35v and offer a couple of XMP profiles to boost performance. There are other modules on the market that will clock higher but for me, these are perfect. The VLP modules clock in at 800Mhz out of the box with a very attractive 9-9-9-24 at a 2T command rate.

Data

MLC SSD prices continue to become more attractive with loads of available performance. At a minimum, running an SSD for your OS disk is a worthy investment. For this build I’m running 3 disks: 1 x 128GB SSD for OS, 1 x 64GB SSD for apps/scratch, 1 x 1TB 7.2K SATA for data. I have an external NAS array for backup and greater data capacity. My OS disk is the highly acclaimed Samsung Pro 840 SATA 6Gb/s that boasts random reads up to 97K IOPS and sequential reads/writes up to 530/390 MB/s, respectively. In other words, pretty damn fast.

For scratch and application files I’m reusing my Kingston SSDnow+ drive and for data I got a new Seagate Constellation ES drive. Any drive worth buying will have a warranty of at least 5 years. Those with 2 tend to wear out faster from my experience.

All hard drives are mounted cleverly behind the motherboard in the SG09. Up to 2 x 3.5” drives or 4 x 2.5” drives. Having done away with optical media a very long time ago, this new PC will not be getting a DVD/BR drive which must be slot loader style in this case anyway.

Graphics

Although the Ivy Bridge CPU now includes a more powerful on-chip graphics processor, it’s better suited to mobile devices or applications with less graphical demand. I opted to continue to run a discrete GPU for high performance graphics. The GTX 670 is a very good high/middle-end card that many benchmarks have pegged near the 680. I chose the 670 over the closely-spec’d 660Ti due to its 256-bit memory architecture which is stunted on the 660 at 192-bit and as a result offers less memory bandwidth. Most of the other specs are almost identical. Near term you would probably not notice much in the way of a severe performance difference but long game, the 670 has longer legs and more headroom. If the price is close, the 670 is the way to go.

Build Summary

As far as sourcing parts goes, it found it interesting that Amazon and NewEgg are really the only reasonable places online to source parts anymore. Sure TigerDirect is out there (compusa) but I found their prices to be generally not competitive. I suppose this does reflect the current state of the custom PC today with fewer builders perhaps but there are still a plethora of available parts. As a long time customer of NewEgg, they appear to be doing quite well and have earned top spot as premier parts destination online. Its return and RMA process still leave something to be desired with restocking fees and customer paid return shipping. If Amazon started stocking more PC parts I‘m pretty sure it could push NewEgg out of the game entirely. The only thing I bought from Amazon for this build was the PSU, consequently. MicroCenter provided the CPU and Pro 840, everything else was cheaper with coupons, shipped free, and tax free from NewEgg.

Windows 8, or…bust?

There’s a lot to like about Windows 8: settings sync between devices, new windows explorer, task manager, file copy/ pause operations, power savings, memory savings, oops reset, and some increased security. There’s a lot not to like about Win8: start screen, devalued search, hot corners, separate UIs to update and manage, a device with no touch negates many of the new improvements. I’ve run every version of Win8 since the first preview release. I’ve had a lot of time to digest it, watch the changes, read the white papers, blogs and get accustomed to the UI(s). I’m sad to say that Win8 will be the first new Windows OS ever that I will not be installing on my new PC. It just doesn’t make sense to me. The majority of the under the hood changes to Win8 deal with optimizing the OS for portability, which is fine, but I’m not overly concerned with less RAM consumption or new sleep modes on my PC. I am sure that Win8 is an amazing product on a proper touch enabled device a la the Surface or Dell XPS 12.

Win8 is a dual-personality OS with 2 distinct UIs each with its own set of applications. Metro apps run sandboxed much like Android or IOS and silently suspend or terminate in the background, much the same way. After a long time forcing myself to adapt, I’ve come to the conclusion that metro just isn’t for me on a desktop PC with a keyboard and mouse. I’m not interested in live tiles or apps from an app store, nor do I want to hover over hotspots or drag windows via mechanisms designed for touch. The new Start screen is essentially an exploded version of the Start menu, but now the search functionality has been handicapped. You can still push the windows key, type cmd and hit enter but if you’re searching for something there are more clicks involved now between the apps, settings, and files sub categories. Windows7 start search is far superior in this regard. Metro (formerly known as) is an interface designed for touch, period. So I resigned myself to “live on the desktop side”, I installed Classic Shell and enabled booting to the desktop which brought back the familiar feel of Win7. Start search is still severely lacking with Classic Shell in place and I’ve had a few application compatibility problems as well. Win8 just doesn’t feel like an upgrade, especially having to make so many concessions to live with it. The user experience overall has gotten worse, not better, in my opinion for non-touch enabled devices. Win8 will likely be the new Vista (which I never had a problem with personally) and Microsoft has some time to watch the market before making their next move. They hedged their bets between tablets and PCs with Win8 and have managed to give each group enough of what they need. If this tablet thing fizzles into a fad, no problem, out comes the full blown desktop, if the PC dies completely metro will be primary. I’ll continue to run Win8 on my laptop, but my new powerhouse desktop will be getting a fresh build of Win7 SP1.

Final Thoughts

Overall I am extremely pleased with my new Ivy Bridge build in the SG09. One of the few aesthetically bothering items is that the ST55-F PSU in the SG09 sits off-center to the left so you can see the PSU logo off behind the case logo. Also the bar of stickers on the right side of the face of the PSU are somewhat distracting. The use of 90-degree sata connectors should be avoided in this case as they can interfere with the closing of the top lid. SATA “6Gb” cables came with the V Gene board but tests show there is no performance difference whatsoever between these and SATA cables from years ago. The 6Gb cables do have locking mechanisms on the ends which are a nice touch.

The OS installation was a marvel all unto its own taking 5 minutes flat from start to finish. Reboots are a literal 10 seconds from click to logon screen. In complete stock form this rig is a beast. I plan to overclock a bit once the TIM has broken in on the CPU but the need to do so is really minimal. It will be for fun more than anything.

From a power consumption perspective, this PC has yet to break 200w of usage, even during peak graphical or CPU intensive operations. For anyone wondering if 550w in a SFX PSU will be enough, yes it will! I have light fixtures that consume more power than this PC! It is also very quiet with all fans operating normally. There is a fan switch in the rear of the case to boost all RPMs to high which does kick up the noise a bit. Between this build and my old rig it’s almost like my office is too quiet now. :-)

Temperatures are so far better than average from what I’ve read, with the CPU idling ~33-34 Celsius. This will get better as the Artic Silver 5 sets in.

I don’t have any other performance benchmark data at this point (other than my 7.9 Windows user experience score) so I’ll just leave it that this thing is FAST and was well worth the upgrade. Faster, quieter, smaller, less power consumed = mission accomplished.

From:

To:

With your FIOS service, Verizon installs an Optical Network Terminal (ONT) which takes the fiber optic signal delivered to your house and converts it to electrical signals that traverse the copper coax in the home. The ONT is responsible for delivering your TV, data, and phone services. With the ONT Verizon provides a battery backup unit (BBU) and I have read that there are different versions of the ONT and BBU so this hack might not work for you. My BBU is the CPL28U12 model. If something goes wrong Verizon will likely not be happy, this will not be my fault, proceed at your own risk (or peril).

The Verizon provided BBU by default protects only the Verizon-provided phone service, which does me no good since I use Ooma. The BBU cover slips off downward with a single screw holding it in on the right above the wire connections.

As you can see this is a fairly sizable battery at 7 amp hours and a shame that it was implemented to only protect one of three possible FIOS services! I suppose they legally erred on the side of caution to keep the phone open in case of emergency but they should still offer to protect the other services if you don’t subscribe to the phone service. Anywho… there is a way to remedy this.

Next to the primary 12v feed there is a wire bundle attached to a pinned module that plugs into the bottom of the BBU. Take note of the wires that plug into the “on battery” and “battery missing” slots, we will switch these in a minute. Mine were blue and grey, yours may not be.

Pull the blue module gently straight down, it will come off in one piece.

The wires are simply pinched down in this module by the clear doors that cover them. No need to strip the wires. There are 3 clear doors that open from the bottom where the wires enter the module. Gently lift up from the areas indicated below. The doors will raise slightly and release the wires in their slots.

With the doors open, switch the “on battery” and “battery missing” wires. The image above shows mine already switched. Reseat the wires and lock the doors down. Make sure they click securely at the bottom from where you first swung them open. Carefully plug the blue module back into the base and you’re done. Now if the ONT looses AC power it will draw from the battery protecting your internet and TV streams. Test to make sure this works.

Props to vzguy71 on broadband reports who discovered this back in 2007. The reason why this works is because in an AC power outage, the ONT now sees the battery as missing but nothing wrong with the AC power (being fed from the battery). Before it could see that it was running off of the battery and would limit the power to the phone service only. This is a very cool and easy change to make use of your FIOS BBU especially if you do not subscribe to Verizon phone service.

Power is and always will be an issue in the datacenter, especially as gear gets denser, faster, and hotter. Power circuits can be sized improperly from lack of due diligence or simply devolve as equipment moves in and out of racks. Having gone through a few datacenter build outs I wanted to share some of my methodology for power sizing and planning at the rack level.

Power analogy

Think of electrical circuits and delivery through copper wires like water through piping. There are 3 main facets that we are concerned with in sizing the power infrastructure: voltage, amperage, and power. Resistance, measured in ohms, is equivalent to the size of the water pipe, but since power cables are provided to us by electricians and vendors, we’re not overly concerned with that aspect here. (Although I should mention that it is important to ensure that the power cables used to connect to equipment power supplies are capable of max PSU draw).  Voltage is equivalent to the pressure (force) of the water flow. Amperage is equivalent to the rate of flow (current) of the water and also helps us determine efficiency. The power of the circuit is determined by multiplying the force (voltage) by rate of flow (amps). Assuming a Power Factor of 1, this equation also determines volt-amps, most commonly rated in Kilovolt-amps (Kva). This of course changes a bit in a three-phase circuit but I’ll get to that later. In either case, increasing the force or current flow will result in more power generated.

Watts or volt-amps (single-phase) = Voltage x Amps

Efficiency

In most cases, when you increase the voltage supplied to a server, storage, or networking device, it becomes more efficient in the way it utilizes power. In other words higher voltage = reduction in current draw. For example, an HP DL380 G6 with a 750 watt power supply can, as rated by HP, draw up to ~8 amps @ 110 volts. If you provide this same server with a 208v power circuit its draw decreases to just over 4 amps. The consumed current continues to decrease as you increase the voltage. These power supplies can go up to 240v reducing the current draw to a nominal 3.6a. That’s a 45% difference in draw between the lowest and highest capable voltages. The advantages of higher voltage in power constrained environments is obvious.

Sizing

So how big do you go? Single phase? Three-phase? 110? 208? 220? An enterprise-class UPS typically has 3 power phases: A, B, and C. The utility power circuits attach here usually as 480v @ 600a mains, two for redundancy. That’s 1200 amps of draw and any number of voltage scenarios that can be built off of the breaker box on the other side of the UPS, assuming it can handle it. Each phase has a maximum current and voltage depending on the size of the breakers installed in the UPS. You can quickly get a general idea on how much current your datacenter is consuming by looking at the load levels of each phase on the UPS, displayed in amps.

Let’s assume that the power infrastructure has been properly built and that the internal UPS breakers are setup so that all available power is being distributed to breaker boxes evenly. We need to figure out what size circuits, called whips, need to be pulled from the breakers to the racks. You can calculate based on averages or peaks. Calculating based on peak loads will make for absolute certain that you have enough available juice to power all of your attached gear if it were running at 100% load. If this is unlikely to ever occur in your environment, you should be ok using the averages.

There are 2 factors you need to figure out to properly size your power circuits: total power required (watts) and how much draw (amps) at a given voltage. As I stated before, this number will vary depending on input voltage. All equipment power supplies are rated by the manufacturers of the devices. From each device you plan to place in a single rack, add the rated draw from each system’s power supply and the required watts of each. You will need to ensure that you can provide enough total power as well as draw.

For example we determined that we need: 6000 watts and 25 amps total with 208v delivery. To meet these demands, assuming two circuits per rack for redundancy, the minimum circuit size we could run would be 208v @ 20a single phase. This yields 4.2Kva per circuit or 8320 watts total delivered to the rack ((208 x 20)x2). If 110/120 power is all that you have available, then you need to re-calculate your equipment requirements because current draw will likely double at this reduced voltage.

If you aren’t completely filling your target racks then don’t forget about providing room for growth. You may only need 6Kva day one, but your rack is only half full. In the example above, the two 208v @ 30a circuits will give us 28% additional headroom. If we plan to double the same type and number of equipment in this rack then these circuits will not be enough. If we plan for a 50% potential increase then we need to provide enough power for 12000 watts of power and 50 amps of draw. To meet this demand we’ll need to increase our circuit size to 30 amps at 208 volts. This will provide 6.2Kva per circuit for a total of 12.4Kva. This is a solid and safe sizing estimate for our new rack.

Three-Phase Power

While extremely efficient and better for redundancy, 3-phase circuits can be more expensive because of the increased power output. This is where we have to remember the overall Kva output of our UPS. If not careful, then we can overrun its capabilities. To calculate the power rating of a 3-phase circuit, start with the formula above, then multiply by 1.732.

Watts or volt-amps (three-phase) = Voltage x Amps x 1.732

Referring back to the scenario above, we need 12000 watts and 50 amps for our new rack. In a single-phase solution we need 30 amp circuits running at 208 volts. If we took those specs and turned them into 3-phase circuits, we would get 10.8Kva per circuit or 21.6Kva for the entire rack, a lot more than we need. Because of the increased output of 3-phase power we could now actually reduce our circuit size to the rack. 208v @ 20a, 3-phase, yields 7.2Kva per circuit or 14.4Kva total which is closer to our calculated estimate. 3-phase power saves us 20 amps of draw from the UPS on this particular rack!

Your particular power requirements may reveal that 3-phase power is unfeasible as it might produce too much power for your given application potentially overwhelming your UPS. I just went through a scenario where single-phase actually made more sense based on current load of the UPS and our power requirements. There are certainly more factors that could be considered but this should give you a good basis for determining power needs in your environment.

References:

Power formulas

Power calculator