I really hope you have met with HPE Primera – correctly pronounced called praɪmɛrə – earlier on, at least here on my blog. Now that more and more information and details are no longer under top secret stamp, it is time to write and talk about it, especially what the configuration options are and how it’s numbers are looking like.
It is available in two types, all flash and hybrid. Be warned that Primera was built to deliver all flash based storage services, it is not meant to be a hybrid system. Furthermore this latter one is not available in many countries and can be purchased if approved by HPE. All flash models are starting with letter “A” while hybrid ones are with “C” – converged flash. C models can have SSD + HDD, SSD only and HDD only, A models can have only SSD as media.
Okay, okay this is not that simple as chassis itself – controller chassis – exists in two sizes, 2U and 4U. In a 2U chassis the controller pairs can be A/C 630/650/670, in the 4U only A/C 650/670 controllers are allowed – two our four. Based on the enclusure size and the number of controllers they start at 24 x 2,5″ drives in a 2U chassis or in a 4U chassis with two controllers and can go up to 48 x 2,5″ drives if in the 4U model, four controllers are installed. Important to note, that controllers cannot be mixed, so in a chassis, only same type of controllers are allowed.
2U controller chassis
As said 24 SFF SSD – or HDD in case of C models – with SAS connectivity for now, later the last 8 slots can have NVMe SSDs.
Simply put all that a 2U chassis can have, but everything doubled. 48 x SFF SAS connected SSD – or HDD in case of C models – and later 16 x NVMe SSD can be added, if the chassis have 4 controllers. If it only has two, the 2U and the 4U chassis is no different.
Then what is the point to have 4U with 2 controllers opposed to a 2U model? The upgrade path. Online and non-disruptive upgrade from 2U to 4U is not really possible, so plan ahead and think about your future needs.
If the above chassis not enough to address capacity and performance needs of yours there are two disk enclosure options, but in reality only one. Size is the same, 2U, but while the SFF model can have 24 x 2,5″ SSD – or HDD in case of C models – the bottom model is only for C models and as its name states 12 x 3,5″ drives are allowed in it.
Above you can see it, but three models are available at this point. The entry model called 630, the mid kit 650 and the top system 670. They can be A or C models as said. The major difference here is that 630 is always a two controller system, while 650 and 670 controllers can work in pairs – so two controller setup – and in four controller configuration. The explanation for this A630 limitation is the architecture of its controller, since it has one Xeon processor and one Primera ASIC. Beefy models have 2 Xeon and 4 ASIC per controller and one role of the ASIC is to communicate with the other controller(s) A630 is limited in this perspective.
Furthermore the achievable IOPS/latency and maximum capacity figures are increasing and scaling towards the larger models.
Hereby note that I will not talk again about the hybrid models as if someone wants still spinning media in a storage go for a 3PAR, Primera systems are not the droids they’re looking for.
So let’s focus first on the maximums which are two different things as they can be measure based upon SSD count or capacity. In case of the smallest model these two interestingly the same, so maximal capacity can be reached exactly with the maximum number of disks – if 1,92TB drives are used. Only two free bays will remain, but this is pretty valid for all models anyway.
As you can see in case of A650/A670 if only 1,92TB – the smallest capacity option of SSD – used in the maximal possible quantity the resulting capacity will not be near the maximal capacity that controllers can handle. Note that the drive counts above will not result a balanced configuration in RAID 10+2 for example, so these numbers are theory numbers.
It is also true that the other end of the scale, so if the minimal number of SSDs are used only, controllers will not deliver their full performance as they will be limited by the media itself. Very important that in vendor marketing materials around the internet and in some presentations the IOPS numbers and latency figures. The IOPS alone is not representative. Let me put here an example:
My car’s performance is 600 BHP. I am not telling the weight. A Ferrari GTC Lusso and a Scania R truck are equal as they are both having 600 BHP. But when you do a race between them they will not perform the same way, since the car weights only at 1740kgs while the truck is around 7500kgs without the trailer.
Also there are some presentations showing the IOPS maximum at 8k write, but reads in 4k, on the same slide. This is misleading and called hero numbers so confuses the reader and makes one to one comparision extremely difficult. A storage system can be junk even if it is the best in write IOPS but zero in read IOPS. And we have not touched latency…..bottom line use your common sense and measure,test,plan.
In the planning phase use your own figures that are specific for your infrastructure.
And here there would be my performance testing, but due to legal restrictions I am not allowed to share it. Basically what I did is I compared the performance of different models scaling up over the course of 8-20-40-80 SSDs. This way I could define the points where no matter how many SSDs you add above that limit, performance will be the same, so only the capacity is increasing. If you are interested in this – I bet you do – contact me directly. While we have zero ability to measure any performance metric, since we have no Primera system at all, I could extract values from the official and “partner only” tool, but for some reason it is top secret. If Primera is the best and last storage customers will ever buy, why these things are confidential….
Anyway I had the luck to be in Kutna Hora and HPE Hungary has an A650 system so saw them already, but I never tested a Primera – and I don’t even want to – as just to put enough load on an array is very difficult task to do. This is valid for designing/planning an array, sizing it properly as it must be known exaclty what the needed workload will do as sequential read/write numbers are looking great, but no one will load the array with IP camera steams in it’s service time 7x24x365.
HPE was not joking when they designed this architecture, which will probably dominate the Tier 0 market in mid term 5-8 years. Especially that performance they give are all based on DECO, so no fancy trick with thin volumes here, and latency diagram is always flat up until the last percent where it saturates. You can expect some articles in this area in the future as raw performance is a must, but not enough in the modern IT anymore as consumption and manageability is dominant now and HPE has some magic in that area as well.