Oracle: What's the X4-4? In-memory monster slams jaws down on 3TB, for starters Oracle can analyse up to 3TB of data in memory with its latest Exalytics In-Memory Machine, the X4-4. Announced at Larry Ellison’s OpenWorld extravaganza, this system features four Xeon E7-8895 v2 processors, up to 3TB of memory, 4.8TB of flash and a paltry 7.2TB of disk storage (6 x 1.2TB 10K 2.5-inch drives). Its X3-4 …
and how scalable I suppose..
Last week HP Vertica was boasting how facebook uses their platform to ingest 35 TB/hour (15GBytes/sec) of new data in a 4PB rolling data set, all without losing more than 20% of query performance for the clients.
Vertica architecture says SSDs don't really help performance since their sequential data rates aren't much more than spinning media.
at least to something bigger than the memory you have. I've seen it happening with Essbase-type products all the time. You get hooked on the first dose and build awesome datamarts that are lighting quick. Then business starts asking for more and you end up hitting the memory barrier.
Solution? Just add more hardware, please. Only problem is that it becomes very, very expensive.
Got a couple of these at work http://www-03.ibm.com/systems/x/hardware/enterprise/x3850x5/specs.html
Apparently there is an interconnect that allows you to single image two chassis together (I assume NUMA)
So that is 4TB of ram & 160cpu threads
There are workloads which you can not run a clustered solution. You need a single scale-up server for those workloads. Scale out clusters won't do. Oracle will release a 64TB server with 32-sockets SPARC M7 cpus next year. It has 8192 threads and 1024 cores.
For instance, to guarantee database transaction integrity over many nodes in a cluster can be really difficult, rollback and all that. The synchronization can be very very very tricky. It is much easier (and faster) to just use one single huge server instead, instead of a cluster. Some companies tries to do this, and after much research and development concludes a huge fast SMP server is cheaper and better than all the R&D.
BTW, there are no larger Linux servers than 8-sockets on the market. So you need to go to Unix or Mainframes for these workloads. SGI Altix and UV servers are clusters, as SGI explains, as they can only handle HPC number crunching workloads serving one single user. Whereas Enterprise SMP servers typically serves thousands of users running business software, typically Unix or Mainframe servers.
Long John Brass:
>> BTW, there are no larger Linux servers than 8-sockets on the market
>Linux will run on IBM Z series; No?
Linux will run on IBM Z series, yes. And Linux will also run on 32 socket IBM P795 AIX server. But both of these servers are not Linux servers. They are Mainframes and Unix servers. And someone has compiled Linux ontop Mainframe and Unix servers. That does not make them Linux servers. I suspect Linux runs bad on them, if we look at the Big Tux Linux experiment, where they compiled Linux onto 64-socket HP Integrity server: ~40% cpu utilization. Which means every other cpu idled, under full load from the Linux kernel.
Again, there are no, and has never been large Linux servers for sale suitable for SMP workloads. SGI and ScaleMP both agrees. SGI Linux cluster Altix and UV2000 are only used for HPC number crunching (serving one single scientist), not for Enterprise workloads (serving thousands users running database or ERP systems). No one runs SGI Linux servers using SAP for instance, there are no SAP benchmarks on Altix or UV2000 SGI clusters.
It's done via the QPI links and what's known as a "glueless" architecture. You effectively jack 2x 4-socket systems together to form 1x 8 socket system. Naturally you need to have the processors that support it; i.e. the E7-8xxx series processors rather than the more common E7-4xxx models. As a result, unless whoever specced/bought them anticipated doing it you'll probably be looking at a rather pricey CPU upgrade to both machines.
IBM also do something similar with one of their blade servers in that you can smush 2x 2-socket blades together to form 1x 4-socket blade, again based on the E7 processor series.
With regards to the Oracle box most of the other vendors now do E7 v2 machines with 6TB RAM. The X4-4 has the same amount of memory slots as those other machines so perhaps Oracle haven't vetted 64GB DIMMs for it yet?
I'm not that convinced x86 has a reliable and robust memory subsystem to support in-memory workloads. I'm sure that will annoy some as thinking they are grown up but think mission critical...maybe life critical in certain cases. Would you want that running on x86? If your Board or C-suite expects you to deliver a solution that doesn't break the bank yet will run 24x7 would your first choice be x86? I would ask why you wouldn't consider a Power8 server? If a 2 socket, 24 core S824 with 2 TB of Ram and using memory compression that could be 4.5 TB equals the performance of a 4 socket 60 core Ivy Bridge server at the same cost, would you not consider it? Compare the security, RAS and virtualization features vs x86, VMware and Linux to see for yourself. Yes, this sounds like a commercial but so was the article. Using the Ford vs Chevy analogy, I'm asking you Chevy owners to consider a Ford. Chevy isn't what makes your business a success. Rather the reliability, security and total cost of ownership does.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
