Reanimate old server cases with HP Common Slot power supplies

Started by socialreger, Aug 09, 2022, 12:27 PM

Previous topic - Next topic

socialregerTopic starter

If you have accumulated unnecessary server cases, do not hurry to scrap them: there is a chance to save €400 - €700 on the purchase of a new platform.



Standard Common Slot power supplies first appeared around 2010 in the Proliant G6 and in most other HP servers of this time. These mass-produced devices are extremely inexpensive on eBay/AliExpress (€30 – €60 apiece). As a rule, power supplies are in good order: cases of their failure or death when turned on are extremely rare. There is no problem with wholesale lots: we buy more than 100 pieces at a time.

Common Slot power supplies are quiet: at low and medium loads, their noise is lost behind the noise of a regular PC. At full load, a small fan emits a characteristic loud screech at 65 dB (however, it is not audible in the data center).
The HP site contains little info on the specifications: it is not possible to find a usage diagram and pinout there, but there is a suspiciously similar Murata Power Solution device from another manufacturer. His site turned up the documents we needed, including detailed manuals for 470W and 1700W power supplies. The results of research by enthusiasts are also available on the network, apparently HP Common Slot was not only to our taste.


What are hаckers talking about?

    The efficiency of HP Common Slot under average load is 93 - 95%.

    To turn on, you need to connect pins 36 (Power Supply Present Signal - short pin) and 37 (12V Standby Output) on the bus through a resistor of about 22 kOhm, so that the PSU decides that it is inserted into the slot, and pin 33 (Power Supply on / off control signal - far right) connect to a protective conductor (ground).

    Up to eight HP Common Slots can work in parallel.

    To turn on the PSU in parallel with the load, no special actions are required: the main load is separate, the Vsb lines are separate (see the table in the Murata Power Solution documentation).

    The PSU has an I2C interface and an analog power indication output (60.15 mV/A).

How to use it?

The effortless option is to buy a ready-made distribution board from AliExpress or another online webstore.

If we don't like ready-made boards, we can make our own. This requires a Wingtat Model 2.54 EDGE SLOT DIP 180° SINGLE LEAF TYPE WITHOUT EAR High Power S-64M-2.54-5 slots power connector.
The firm is quick to respond to inquiries and send orders for 50+ connectors for around $3 each. They are available retail on AliExpress. If you plan to use a power supply unit of all 1200 W, it is better to order a board with thicker copper. It is worth soldering the connector on it with a hairdryer and thin solder with flux: this is the fastest way, an ordinary soldering iron does not normally take such boards.


How does it look on the server?

    We often use Supermicro universal cases and many of them no longer have out-of-service or upgraded original power supplies for new motherboards. It is not advisable to look for native PSUs, since they are expensive and just as well-worn.

In this case, the native power supply became unusable and was replaced with a universal 750 W HP Common Slot. We cut off the excess and fixed the block in the case with double-sided mounting tape. To convert 12 V to the voltage necessary for ATX24 operation, native PicoPSUs or their Chinese counterparts are used.
The photo shows a Supermicro H11DSi server with two AMD Epyc processors and a consumption of about 400 watts. He categorically refused to work with Chinese counterparts, and his native PicoPSU did not fit in height and was turned on through an adapter. The power supply of the processors is switched on directly, this does not interfere with anything.


Why is that needed?

    We save about €600-€700 on new platforms. This gives us to keep the cost of renting dedicated servers at the market level, plus we do not throw away old cases with all the fans and backplanes, so as not to pollute the environment in vain.

In terms of reliability: usually the power supply fails the first time you turn it on. We have a special test stand for this with a 5 A category A automatic machine. If the PSU turns on, it will continue to work for years. The failure of a power supply unit of this class during operation in a data center on our fleet of some thousand machines occurs approximately once every six months, as a rule, during equipment upgrades. PicoPSU in ordinary servers practically never fail.

Let's try to use power supplies in pairs and assemble a test bench based on our server for this.


Does it really work?

In the video, we will run Linpack on a server with both power supplies connected and will disconnect them from the net one by one under load, turning them back on after a while. We will make sure that the voltage remains normal and the server does not reboot under the stress test. The load according to the server is about 400 W under a stress test on processors, which means that about 500 W is leaving the network.


Why is it necessary to use two PSUs?

Many of us think that redundancy is needed for fault tolerance: if one power supply fails during the operation of a server in a data center, the second one will provide work. This opinion is partly correct, but there are other nuances.

    Usually, when the power supply fails, it knocks out the machine in the remote control or in the shield (due to a short circuit or overload). If both units are included in one remote control, redundancy will not help. Data centers are often designed so that two power beams that do not intersect each other come to the rack.
They have dissimilar distribution cabinets, they go to different UPSs, and these UPSs are connected to dissimilar transformers and diesels. All this switches with some delay, and the data center ensures that power will always be on one of the beams, but not on both at the same time.

It is principal that the server has two power supplies, or you will have to use a rack ATS - a rather expensive device with a solid-state fast relay that can shortly and imperceptibly to the load switch between phase-out-of-phase inputs.

Results

If empty server cases have formed in your household, do not hurry to throw them away: old hardware is easy to put in order. The solution described in the topic will allow you to organize an assembly with guaranteed or redundant power quickly and without significant costs.
  •  

Drupas

Ability to bring the server to maintenance without interrupting the work of services must always be. Whether migrating services to other capacities, or taking capacities offline through load redistribution is a matter of course.
I have a part of the super micro generally in 1bps performance, it died and died, the cluster is redistributed automatically, it just blinked with alerts. But if you have a dedicated 1 server - such functionality is no longer applicable and service work on a hot one becomes necessary.

PS. from personal operating experience, 96% of replacements are dead screws, 2% - bp, 2% - i / o in the storage system. 1% - other. Video card was dead. I almost never met mother boards, possessors, memory, who died in the process. Memory can be broken, but it is rejected at the stage of pre-operational tests, and not during the process.
  •