I am trying to determine the relative difference in performance between the following server platforms that will be running an Oracle 10G RDBMS application supporting 3,000+ concurrent users:
1. Sun M4000 SPARC VI 2 CPU (4 core, 2 threads per core), 5MB L2 cache
2. Sun X4600 AMD Opteron 8222 4 CPU (8 core, 1 thread per core), 1MB L2 cache
Sun tells us that the M4000 with 2 SPARC VI cpu's will "run circles" around the X4600, even when equipped with 8x8222 dual core chips. I can't find any data to backup up this claim.
The current configs on each machine are driven by the available budget and Oracle per CPU/core licensing.
I have located some relevant data on the Sun M4000 but have yet to find anything for OTLP workloads that has been published relative to the X4600.
Any thoughts, insights? In order to use the X4600 I need a compelling reason to do so - and improved performance would be the key here. I'm not looking for a discussion relative to the pro'/con's of Sun's proprietary SPARC versus the AMD Opteron. The customer in this case is a data center shop running big Iron and typically does not purchase lower end servers such as the X4600. However, until last week the X4600 was the only option until Sun lopped $30k off of the price of the M4000.
The operating platform in either case will be Sun Solaris 5.10 (SPARC or X86) backed by an EMC Enterprise SAN. Will likely run the database on raw volumes.
A follow-up to this story and the result was a pretty big surprise. Sun provided under non-disclosure benchmark comparisons between a variety of their mid range servers (AMD's, Xeons's, and Sparc VI)
I was skeptical but what it showed was that performance was fairly equal until the system was stressed. Under those conditions the Fujitsu SPARC VI class (M4000 in our case) seperated from the pack.
We have put the system in production and the 2-5x performance guarantee over our 6800 was easily met or exceeded. I have run a variety of workloads and some TPC benchmarks. Under worst-case sterss the TPC #'s were degraded by only 30% or so.
I was initially worried about the 2xdual core CPUs. However, what I find is that we are geting about 8CPU performance out of this box due to the 2 vertical threads per core. mpstat shows a good balance of workload across all 8 cpu cores/threads. We've hit a peak of 6,000 I/O's per sec.
The M4000 is an excellent mid-range RDBMS server. Originally I really thought the AMD would be the better choice but the Fujutsu designed enterprise class machine is exceptional and highly discounted (we purchased 2 for $60k but that was for a large gov't entity and Sun wants the Fujitsu machines in the large data centers). I'm sure for different types of workloads the X4600 could be as good or better, but for our descrete OLTP and heavy reporting workloads I am very happy with the M4000.
What about quad cores? Barcelona gives you four cores in the same socket and Shanghai is just around the corner.
The problem with TPC-C benchmarks is that they are really an exercise in I/O subsystems. The typical TPC-C test has hundreds, if not thousands of spindles. Not a realistic configuration for any customers that I know.
The best benchmark is your app, your database and your data.
Yes, that was the reason for my follow-up. The tests were run on our configuration, not a TPC test platform. The TPC tests were run while the application was placed under an application specific load scnerio. My only point was to followup with my experience relative to scalability on our OLTP application. The M4000 handles stressful scenerios very well.
This post isn't meant to be an AMD versus Sun SPARC comparison, but just a follow-up on what we ended up doing with our buy. The new Sun AMD's are quad, their Xeon Quads (X4450?) also had good benchmark results. Everyone has quads now, Sun's SPARC VII is a quad. The SPARC VI dual core chips on our M4000 behave like a quad when running our Oracle OLTP application, not a dual core. For database apps the SPARC VII quad-core will probably behave like an 8-way, assuming they still have the 2 threads per core.
Next time I will still look at the AMD servers and it will be likely that Sun won't be giving away one of their M series boxes and I'll buy an AMD. Sun has done a nice job with their AMD designed boxes and the Fujitsu designed M class boxes are a major improvement over their Enterprise boxes for workhorse OLTP loads.
I'm not a Sun evangellist, the customer runs Sun SPARC only and due to budget constraints I was pushing the AMD. Sun offered to give us the box for $30k but we refused until we could see some benchmark data and verify those results with our application. We were given the benchmarks under an NDA. Our real-world experience tracked the benchmark results very closely, a result that I was not expecting.
AMD has always been good and Sun has made great strides with these new SPARC chips.