Friday, March 29, 2013

Honey, I Positively Pressurized the Hot Aisle!



Image 1

In this post, I try to reconstruct a situation I saw during a field visit to a friend's colo data center: An Unintentional creation of positive pressure in a hot aisle resulting in the overheating of idle servers. (Image 1)  With all good intentions, my friend (the owner) had invested in creating a hot aisle containment system for one of his clients.  


Image 2



For one reason or another, (not consulting with an engineer) they improvised by creating a "hot pod" that contained the hot aisle of two rows without a top chimney (connection to a return source to the CRAC units). Image 2. The end result was a hi tech-looking pod that looked like a duck, walked like a duck.... but it didn't quack like a duck. 

When my friend described the situation to me, my conclusion was that he somehow managed to positively pressurize the hot aisle; but how? The answers can be shown using CFD simulation.  The model was created by my colleague Eric Fournier (an expert in CFD modeling!!) for purposes of this blog only.  We exclusively use 6Sigma Room from Future Facilities for all our CFD modeling. We assumed densities across the rows as that information was not shared with me. The largest density 10 kW/rack, the smallest 3 kW/rack.

Complaints were flowing in from the leasing client that the face of the idle servers are too warm (standing in the cold aisle).  The facility operator also noticed that they are constantly having to over-cool the cold aisles to compensate for rising temperatures along the front of the idle racks (total waste of energy) and avoid breaching of Service Level Agreements (SLA's). Image 3 shows the same effect. 


Image 3


The concept of containment and releasing heat works when you create a differential pressure.  In household chimney or furnace design, you have to create a draft using a long vertical stack with an intake at the low point and vent at the high point. This is a natural vent concept. A pressurized vent is similar but you have to physically connect it to a negative pressure source (return fan).  In a contained system, you need that physical connection and a negative pressure source: a duct, return plenum, direct connection to the CRAC unit return fans.  All these components are missing in this installation.






Video 1.  Pressure Across Isometric View

Video 1 above animates variations in pressure as it slices through the horizontal planes from the ground and up to the high bay ceiling. The first observation is that pressure is all over the place within the contained pod as compared to the adjacent open hot aisles.  Most alarming is the positive pressure at the idle servers at the end of the rows, which is the same as Image 1. (pause the video to see various slices). In essence, hot air is pushing right into the back of the idle servers.  

In the open hot aisles, pressure is more "even" with higher pressure logically in the center and towards the top of the hot aisle as the air exits the servers and leaves the aisle horizontally and vertically back to the CRAC unit returns; hence minimal re-circulation back into the servers. 



Shown below are two video which animate variations in temperature through a section and an isometric view. The CFD shows a plume veering to the left side of the pod perhaps as a result of the location of one the CRAH units.  This could be adding to the effect of the inefficiency of the air leaving the top of the pod and putting higher pressure on the idle servers on the left row. If this pod had a chimney, air would be leaving evenly in the center of the hot aisle.




In my opinion, these videos doesn't really show an edge for the contained pod versus the open hot aisles; the hot air plume is just as efficient  leaving the hot aisle for the non-contained (slightly better if anything) as it does for the contained.  As a reminder, this is modeled by assuming rack densities all over the place. If higher densities are involved (>10 kW) the results would be different.  







Lessons learned:
  • Pressure can play a number on you and you need to understand how to manage it. 
  • Properly designed containment systems work great; it is well worth to involve a qualified person to design one. 
  • Improvising designs is dangerous. A poor design can lead to breaching of Service Level Agreements (SLA's) and often ends up with a client packing up their servers and walking away to a more reliable site. 


4 comments:

  1. Hi Ramzi, I really enjoyed this post which perfectly highlights the benefits of simulation in mitigating the risk of hotspots in data centers. I’m delighted that you chose to use our 6SigmaDC software to simulate the problem – investigating and modelling a performance issue like this so that you can see what’s happening is exactly what we built it for!
    Like you we’ve seen a number of improvised containment systems, with varying levels of sophistication in design and construction, and success! I think that many data center operators are attracted to the concept of containment by the sheer simplicity of the idea – if I separate hot and cool airflows the system will be more efficient. But as you’ve illustrated, sometimes what seems like a good idea conceptually can actually create unforeseen operational problems in practice. Such issues can be difficult to clearly understand without the use of CFD to simulate temperature build-up and airflow.
    Standard data center architecture is effectively an open environment for airflows, with hot and cool air mixing within the facility. Although airflow mixing is inefficient, if there is cool air ‘floating around’ the data center there is a good chance (where chance is probably the operative word) that any re-circulated air won’t be that hot. You’ve rightly concluded that an uncontained data center might actually be more effective than a poorly designed contained one.
    As a specialist in CFD simulation for data centers, we advocate the use of modelling for the whole lifecycle, from cradle to grave, including design, upgrades, additions and equipment moves. The obvious benefit is that modelling allows you to answer the question “what happens if” before any work is carried out in the data center. Using simulation, you cannot only see the overall effect of containing the hot aisle in this case, but also what happens when servers are idling or off and creating a recirculation issue.
    Whilst airflow segregation is notionally a good idea as it can increase the efficiency of the cooling system, containment can disrupt airflow for good or for bad depending on the design. Ironically, deliberately preventing the mixing of the hot exhaust air from the cool supply air means that if there is recirculation there is a good chance that the air will be hot, maybe hot enough to potentially pose a much larger threat to the IT equipment we are trying to protect.
    Your observations and conclusions about air pressure in the data center are interesting. Did you model the effect of positively pressurizing the cold aisle or additional flow controls to try to balance the airflow?

    ReplyDelete
    Replies
    1. Mark, thank you for your comments and your insight. Problems in the cold aisle would arise if it was under negative pressure, which did not occur here. I'll have more CFD blogs in the future and will keep you informed. Each day is a school day as one of folks commented earlier! +

      Delete
  2. This shift to cloud computing is really gaining speed, yet the sales of data cabinets and other data center essentials are increasing. It's an interesting contradiction.

    ReplyDelete
  3. It is important to measure the right temperature in the server area to avoid overheat and hanging.

    ReplyDelete

Note: Only a member of this blog may post a comment.