There is a lot of information in the public domain that aids in making educated guesses as to what Big News(tm) will emerge from Oracle OpenWorld 2012.
I recently posted a sneak peek based on information disclosed at Enkitec’s E4 conference. I’d like to reiterate that I do not disclose Oracle confidential information nor have I ever. I signed a binding agreement with Oracle back in 2007 and I honor that agreement. The topic of non-disclosure is fresh on my mind after reading today about the former Navy Seal in hot water for what actually reads just like the legalese I subjected myself to back in 2007. I had no idea service members signed that sort of thing these days. But I digress.
Andy Colvin just posted some of his educated guesses which triggered this post. He mentioned an Exadata X3-8. That really caught my eye and I’d like to explain why. Both the Sun Server X2-8 and Exadata X2-8 (same stuff different SKU) are based on Xeon E7 processors. Those servers, formerly known as the Sun Server x4800 are glueless QPI-midplane 5U boxes that “scale” to 8 sockets. I quoted the word scale because while they do scale from the standpoint of having more sockets physically present in the server they do so at the cost of introducing 2-hop NUMA remote memory and unless one takes the steps to implement application affinity one subjects SGA memory references (including critical code protected by mutual exclusion primitives) to 2-hop latency. Not fun. But that’s not the point I wanted to make about the notion of a X3-8.
I’ll make this short. Sandy Bridge Xeon processors (E5-2600, E5-4600) only have 2-QPI links and they do not support glueless 8-socket designs. Unless Sun emerged from the grave to radically change the pre-Oracle acquisition x4800 design to implement “glue” (ala HP DL980 PRIMA or IBM eX5) the X3-8 will not include a CPU refresh. I’m willing to admit I could be totally wrong with that and thus the word guesses appears in the title of this blog entry. I have no guesses what warrants a moniker change (from X2-8 to X3-8) if not a CPU refresh. We’ll all know soon enough.
As Andy points out the X3-2L is a storage server and the obvious go-to for Exadata Storage Server cells. This model supports 24 10K RPM drives of 900GB capacity. Even though 10K RPM drives scan data slower than 15K RPM drives, large transfer requests cover the evils. Additionaly, we all know that Exadata is not designed for OLTP/ERP so no worries on the damage to seek times. That said, the server should come with an LSI controller refresh and since these things are standard, you’ll likely hear that each cell will support 3GB/s scan rates from HDD alone. Exadata scans flash and HDD concurrently so the aggregate should get a boost as well due to the refresh in Flash technology. However, all that added scan bandwidth will not improve complex query thoughput. None of these incremental improvements address the fundamental asymmetry that throttles physical I/O in Exadata. If you’d like to elevate your understanding of such concepts from faith to fact I urge visiting this post and viewing the video presentations referenced therein.
Storage Indexes and Buffering?
The X2 Exadata product offered 12 600 GB drives so in a gross-to-gross comparison the 24 x 900GB prediction would yield an increase from 7200GB to 21,600GB per cell or about 3-fold. That has ramifications for Exadata Storage Index technology. Remember, 8 columns in every 1MB storage needs to be mapped so a 3x increase in HDD capacity means a 3x increase in the memory footprint to map it. I know the DRAM ratio to HDD mapping requirement for Storage Index technology but will I not disclose. I just wanted to share the fact about mapping because most people–including those who rely on it the most–do not really understand Storage Index technology. There will likely be a boost in DRAM regardless of the reason. Remember, cellsrv must pin down every byte of send data until it reaches the receiver. Buffering takes resources.
Andy speculates the cells will get the E5-2660 Xeon parts. I agree, even though they are slower than the predecessor (Xeon 5675) on a per-core basis once you have more than 4 active cores in a socket. The E5-2600 family uses the 2nd QPI link for added socket-to-socket bandwidth. That along with having more cores is a good thing for the type of processing done in an Exadata cell. But that’s not my point. The Xeon 5675 of the X2 family is a 95W part. Oracle will likely stay in that TDP because they are adding more power draw by pure virtue of the added spindles. The E5-2660 is a 95W part. That takes me to host CPU.
Andy predicts top-bin E5-2600 SKUs for the X3-2 hosts. I don’t share that prediction. The E5-2690 is a 135W part. There are 16 of them in a full rack. Exadata is already an unbelievably power-hungry product. I honestly don’t think they can afford adding (at least) another 640W to the database grid alone. I’m predicting 115W parts (e.g., the E5-2670.)
In spite of the fact that Oracle did not ship the 3,000 Exadata units predicted for Oracle’s last fiscal year–a prediction made at last years’ OpenWorld–Oracle has, in their mind, claimed victory over all others where Exadata is concerned. At least that’s how I see it. As long as people believe it–it must be true.
So, I actually predict some of the biggest “news” will not be in the Exadata space. Mr. Ellison will use this time to thump chest about the SPARC T5 because everyone knows Oracle is up to the challenge of taking on Intel in CPU technology. There will also be no end to the bravado over ZFS Storage Appliance (a.k.a., S7000, Amber Roads) and Pillar (no not the law suit). See, if everyone starts to believe these non-competitive, conventional Sun/Pillar storage products are better than the market leaders’ products (EMC, and, yes, NetApp) then, surely, that will become a fact. At least that’s what I’ve heard.
Finally, we may hear more words of wisdom about “cloud” than mere mortals can tolerate. However, if my prediction is wrong in that regard, we’ll hear bashing of the fine people and products of Workday, Salesforce.com and every other company that has a product to sell into an enterprise IT shop.
I could be wrong on any, or all, of these predictions.
Filed under: oracle