Today, one of the biggest barriers to large-scale SAP HANA environments is hardware. It’s true that HANA is expensive but it also provides massive value that justifies the cost for the right scenarios.
But for large environments, HANA hardware is huge. Let’s take a typical 8TB IBM HANA appliance (remember you get great compression so this is equivalent to 40TB of Oracle). But still you need 16x512GB IBM nodes, each of which has a 4U server and 4U storage node. That’s a massive 128U or 4 32U server racks. You need to locate, power and cool this much equipment:
You can buy 1TB EX5 nodes now from IBM but they are really just 2 512GB nodes stitched together with a special connection, so it is not any more dense. HP, Dell, Fujitsu, Hitachi and Cisco all make HANA hardware too, but it’s all roughly the same size.
The reason for this is because SAP HANA is highly optimised for the massively fast Intel Westmere-EX platform, which has really fast memory. The smaller Westmere-EP blade servers don’t have the grunt to run SAP HANA in an optimal way.
Enter Intel Sandy Bridge-EP
There won’t be a new EX platform until some time late 2012 or early 2013 – Intel is in no hurry because the EX platform is highly profitable and fast. But Sandy Bridge EP is out now and it provides some very interesting characteristics.
Westmere-EX. 40 2.4GHz cores, 512GB (4xDDR1333), 2x 6.4GT/sec QPI
Sandy Bridge-EP, 32 2.9GHz cores, 1TB (4xDDR1333), 2x 8GT/sec QPI
What this doesn’t show is the details of this great Real World Technology blog, which shows that for I/O intensive requirements like SAP HANA, Sandy Bridge-EP massively outperforms Westmere-EX. It has faster interconnects and lower memory, plus 50-70% better core performance.
Based on this it should be possible to run 1TB RAM in a (much smaller) single node with the same performance as Westmere-EX. Interested yet?
What will these systems look like?
Well Westmere-EX can take 1.5TB RAM in 48 sockets but you get a performance hit. As a result I think the right system will be a 32 RAM socket 1TB 32-core blade. Based on what I have seen so far in the blade server market, this means 8x1TB blades in a 10U blade chassis – plus roughly the same in shared storage. We are down from 128U to 20U – over 6x more dense.
Plus the cost reduction is extreme because Sandy Bridge-EP is much cheaper. The expensive item in a 1TB HANA node is the Fusion-IO log storage.
Does this pose a problem for the Intel sales engine?
In my opinion, it’s a big yes. SAP HANA is well suited to scale-out platforms and the highly profitable for Intel. With Sandy Bridge-EP you can use fewer, cheaper, less profitable CPUs for the same effect. HANA simply doesn’t need the scale-up that Westmere-EX requires.
What’s more this is big for a customer. A typical 1TB node currently costs around $200k. With Sandy Bridge-EP and the new FusionIO IODrive2 cards, you are looking at bringing this down below the $100k mark.
Interesting times indeed