Kevin Closson's Blog: Platforms, Databases and Storage

It has been quite a while since my last Exadata-related post. Since I spend all my time, every working day, on Exadata performance work this blogging dry-spell should seem quite strange to readers of this blog. However, for a while it seemed to me as though I was saturating the websphere on the topic and Exadata is certainly more than a sort of  Kevin’s Dog and Pony Show. It was time to let other content filter up on the Google search results. Now, having said that, there have been times I’ve wished I had continued to saturate the namespace on the topic because of some of the totally erroneous content I’ve seen on the Web.

Most of the erroneous content is low-balling Exadata with FUD, but a surprisingly sad amount of content that over-hypes Exadata exists as well. Both types of erroneous content are disheartening to me given my profession. In actuality, the hype content is more disheartening to me than the FUD. I understand the motivation behind FUD, however, I cannot understand the need to make a good thing out to be better than it is with hype. Exadata is, after all, a machine with limits folks. All machines have limits. That’s why Exadata comes in different size configurations  for heaven’s sake! OK, enough of that.

FUD or Hype? Neither, Thank You Very Much!
Both the FUD-slinging folks and the folks spewing the ueber-light-speed, anti-matter-powered warp-drive throughput claims have something in common—they don’t understand the technology.  That is quickly changing though. Web content is popping up from sources I know and trust. Sources outside the walls of Oracle as well. In fact, two newly accepted co-members of the OakTable Network have started blogging about their Exadata systems. Kerry Osborne and Frits Hoogland have been posting about Exadata lately (e.g., Kerry Osborne on Exadata Storage Indexes).

I’d like to draw attention to Frits Hoogland’s investigation into Exadata. Frits is embarking on a series that starts with baseline table scan performance on a half-rack Exadata configuration that employs none of the performance features of Exadata (e.g., storage offload processing disabled). His approach is to then enable Exadata features and show the benefit while giving credit to which specific aspect of Exadata is responsible for the improved throughput. The baseline test in Frits’ series is achieved by disabling both Exadata cell offload processing and Parallel Query Option! To that end, the scan is being driven by a single foreground process executing on one of the 32 Intel Xeon 5500 (Nehalem EP) cores in his half-rack Database Machine.

Frits cited throughput numbers but left out what I believe is a critical detail about the baseline result—where was the bottleneck?

In Frits’ test, a single foreground process drives the non-offloaded scan at roughly 157MB/s. Why not 1,570MB/s (I’ve heard everything Exadata is supposed to be 10x)? A quick read of any Exadata datasheet will suggest that a half-rack Version 2 Exadata configuration offers up to 25GB/s scan throughput (when scanning both HDD and FLASH storage assets concurrently). So, why not 25 GB/s? The answer is that the flow of data has to go somewhere.

In Frits’ particular baseline case the data is flowing from cells via iDB (RDS IB) into heap-buffered PGA in a single foreground executing on a single core on a single Nehalem EP processor. Along with that data flow is the CPU cost paid by the foreground process in its marshalling all the I/O (communicating with Exadata via the intelligent storage layer) which means interacting with cells to request the ASM extents as per its mapping of the table segments to ASM extents (in the ASM extent map). Also, the particular query being tested by Frits performs a count(*) and predicates on a column. To that end, a single core in that single Nehalem EP socket is touching every row in every block for predicate evaluation. With all that going on, one should not expect more than 157MB/s to flow through a single Xeon 5500 core. That is a lot of code execution.

What Is My Point?
The point is that all systems have bottlenecks somewhere. In this case, Frits is creating a synthetic CPU bottleneck as a baseline in a series of tests. The only reason I’m blogging the point is that Frits didn’t identify the bottleneck in that particular test. I’d hate to see the FUD-slingers suggest that a half-rack Version 2 Exadata configuration bottlenecks at 157 MB/s for disk throughput related reasons about as badly as I’d hate to see the hype-spewing-light-speed-anti-matter-warp rah-rah folks suggest that this test could scale up without bounds. I mean to say that I would hate to see someone blindly project how Frits’ baseline test would scale with concurrent invocations. After all, there are 8 cores, 16 threads on each host in the Version 2 Database Machine and therefore 32/64 in a half rack (there are 4 hosts). Surely Frits could invoke 32 or 64 sessions each performing this query without exhibiting any bottlenecks, right? Indeed, 157 MB/s by 64 sessions is about 10 GB/s which fits within the datasheet claims. And, indeed, since the memory bandwidth in this configuration is about 19 GB/s into each Nehalem EP socket there must surely be no reason this query wouldn’t scale linearly, right? The answer is I don’t have the answer. I haven’t tested it. What I would not advise, however, is dividing maximum theoretical arbitrary bandwidth figures (e.g., the 25GB/s scan bandwidth offered by a half-rack) by a measured application throughput requirement  (e.g., Frits’ 157 MB/s) and claim victory just because the math happens to work out in your favor. That would be junk science.

Frits is not blogging junk science. I recommend following this fellow OakTable member to see where it goes.