IBM readies Exadata killer
Big Blue can swing both ways, too
Exclusive According to sources familiar with the company's plans, IBM is getting ready to launch its own database clustering box for online transaction processing (OLTP) and data warehousing.
The machine, which is apparently going to be called DB2 Pure Scale, is obviously meant to blunt the attack of the Exadata 2 box cluster that Oracle and soon-to-be acquisition Sun Microsystems launched in mid-September.
This machine is based on Sun's x64 rack servers and flash and disk storage, as well as Oracle's Real Application Clusters extensions to its 11g database and its Exadata storage software. It is a precursor to an expected Exadata variant due from Oracle on October 14, presumably based on Sparc T2+ servers and maybe a souped up flash array called the F5100, which crams 80 flash drives into a 1U chassis, delivering 4 TB of flash capacity and 1 million I/O operations per second of storage throughput.
IBM doesn't want to let Sun and Oracle do all the talking, of course. This is why DB2 Pure Scale is being launched - perhaps this week ahead of Oracle's OpenWorld show, which starts on October 11 - and almost certainly ahead of whatever Sparc announcements Oracle and Sun make on October 14. IBM could, however, keep its powder dry and do a launch of its DB2 clusters after Oracle is done talking and after it has spoken to Wall Street on October 15 about how its third quarter worked out. Generally, IBM likes to have big announcements out the door just ahead of its Wall Street pitch, so it seems reasonable to expect a launch sometime this week or early next week.
While the details are sketchy, DB2 Pure Scale is apparently a clustered implementation of IBM's DB2 database for Unix, Windows, and Linux platforms. IBM has had parallel implementations of DB2 for years, but DB2 Pure Scale apparently has a slightly different twist.
Remember DB2 Parallel Edition for IBM's RS/6000 SP PowerParallel machines, which were the basis of the chess-playing Deep Blue boxes, back in the mid-and-late 1990s? Or how about DB2 Multisystem for the OS/400-based midrange Power boxes, which was released in 1996 and which does essentially what Oracle RAC does? Or maybe you remember DB2 Integrated Cluster Environment for Linux, which came out in 2005 and which offered clustering for capacity and high availability for IBM's x86 and x64 servers? DB2 and Parallel Sysplex have been around clustering for capacity and HA since the 1990s as well.
According to the sources who spoke to El Reg, DB2 Pure Scale will run on Power-based servers with the AIX operating system and will make use of an InfiniBand interconnect to link server and storage nodes together in a cluster. The DB2 Pure Scale setup will use the Remote Direct Memory Access (RDMA) features of InfiniBand to give nodes access to each other's data for processing. The impending IBM cluster will also have a designated server in a cluster that functions much as a head node in a supercomputing cluster. In this case, it will manage the locking of database fields as transactions and queries are processed and the locking and unlocking of memory in all of the nodes in the cluster.
Apparently, other clustering technologies - this must mean Oracle 11g RAC - have a lot of chatter back and forth as transactions run on the cluster, with each server acting like its own traffic cop. But DB2 Pure Scale just has one traffic cop for the whole cluster (presumably with a hot spare or two) that works like a giant memory controller and database lock. How this can result in scalability without bogging down the central database and memory locking server is a bit of a mystery. But that's why software engineers and math majors still get good paychecks, and I look forward to learning more about this.
It looks like IBM will pitch DB2 Pure Scale in much the same way as Oracle is pitching Exadata 2: It swings both ways. Meaning, it can do OLTP or data warehousing, and provide high availability because it is an inherently clustered environment. DB2 applications will apparently run on it unmodified, just as Oracle is claiming Oracle applications can be with Exadata clusters. ®
DB2 PureScale is an Oracle RAC Killer.
RAC is something that merits being killed. Exadata is just a headline for Larry-Needs-To-Tell-Something-Good-Around-Snoracle.
I mean, DB2 PS is almost linear scalability vs. very poor scalability, application transparency vs. cluster-aware applications with constant architecture tuning and code modifications, automatic workload balancing, instantaneous recovery from node failure.... and much more.
And yes, it runs on Power Systems
Don't compare with Exadata. , which is just an abortion ("A waste of time", by rch)
Sun+ParAccel Exadata-killer: Where did it go?
Last year, perennial TPC under-achiever Oracle sticks a low-tech query filter and gobs of DRAM in front of the disk-pool (IBM circa 1990's, EMC circa 1980's), adds Infiniband (IBM circa 2002), and clusters it up into something dubbed "Exadata"...and pronounces it "revolutionary".
Meanwhile, Exadata performance on all but the simplest data warehouse schema turns out to be only marginally better than Oracle's (famously poor) data-warehousing performance WITHOUT Exadata. Meanwhile, Oracle's "bizarre" TPC configuration tweaks have been noticed...but not explained...
Now this year, Oracle switches from (hp)Intel to (sun)Intel servers, adds some Flash, and once again it's "revolutionary"! To prove it, Oracle promises a world-beating TPC benchmark to show how magical these Oracle/Sun synergies really will be.
But strangely and for no apparent reason, Sun's ALREADY world-beating June 2009 TPC result on X4540 platform and ParAccel Analytic Database has just got "disappeared" from the TPC website. Sun's 30TB result was not only the world's fastest, it was also the world's largest, and (among large systems) the world's best cost/performance -- 1/16th the cost of #2. The Sun/ParAccel results vanished just the week before Oracle's premature mediajaculation about a "world's best" TPC on Exadata2, leaving (get this...) an Oracle/HP solution as the lone result in that category.
Ostensibly the Sun result was "withdrawn" on protest of an anonymous TPC member, even though Sun's Full Disclosure report went through two revs, and neither Sun nor ParAccel has ever had a problem with their numerous previous TPC results.
Maybe the amazing Sun-ParAccel result -- over 1,000,000 QpH on a 30TB warehouse at only $2.86 per QpH -- was "withdrawn" because it would have made the upcoming Exadata2 results look sick. Consider: Oracle's previous result on a 30TB warehouse was 150,000 QpH at $47 per QpH. That's only 15% of the performance of Sun/ParAccel, and at 16x higher cost. Furthermore...even with the advantage of a much smaller 1TB warehouse -- and with the further advantage of having 2TB of DRAM to hold it all, the best that Exadata1 could manage was $5.42 per-query-per-hour.
So...the Sun/ParAccel "world's best" result vanishes just before the "new world's best" result is supposed to come out on Sun/Oracle's Exadata2. Anybody smell fish?
If EU regulators REALLY want to know if Oracle's buying Sun is anti-competitive, they should find out what happened to the Sun/ParAccel solution.
When/if the Oracle/Sun TPC results do come out, the important number will be COST/performance, not absolute performance. Any major vendor today can throw enough "brute-force" (aka hardware) at a problem to overcome even the crappiest software. Good architecture on the other hand is demonstrated via the best COST/performance and performance/energy ratios. By this measure, the Exadata1 been a total failure.
I guess we'll wait for the numbers on Exadata2, and then we'll wonder how it might perform against the competition that got disappeared.
(for anyone interested in digging in to the story, the disappeared Sun results have been de-linked at the TPC website, but are still available here):
(for anyong interested in discounting ParAccel vs. Oracle based on it's columnar approach, check here first):
It's all about balance
In my experience with both IBM,SUN and HP. It's all about balance.
On Sun I get great parallel throughput but to be honest crap single threaded performance, on IBM I get great single threaded performance and often bad parallel performance especially if I have a number of heavy single threaded jobs running.
HP often falls in between, with the best and worst of both worlds.
Now solaris on x86, has in most cases given me the best bang per buck, good single threaded performance and pretty good parallel performance, specially with 6core (AMD) or multi thread (intel).
In specialist cases (i.e. OLAP) I generally go for and IBM solution or for web type transactional loads I've gone for Sun