Category: Hardware

Robert Davis wants to find information on soft-NUMA in his SQL Server instance:

So having read up on automatic soft-NUMA, I was eager to see what it did with my main production servers when I upgraded them. My main pair of production servers (they are paired into an Availability Group) have 4 NUMA nodes with 16 physical cores per node and hyperthreading for a total of 32 logical cores per node with 1.5 TB of RAM. Obviously, we are using core-based Enterprise Edition for these servers. I thought I knew what automatic soft-NUMA would do, and wanted to confirm if my expectations were right.

Read on, but it looks like there’s a “to be continued…” here.

Joe Chang throws down the gauntlet:

Naturally, the database and application should be architected together with the SQL Server NUMA tuning options to support good scaling on multi-socket NUMA systems. If we neglected this in the original design, I am sure many DBA-developers know how well such a suggestion would be received by management.

Is there another option? Well yes. Get rid of the NUMA latency issue with a non-NUMA system. Such a system has a single processor socket, hence one memory node. Before anyone scoffs, the one socket is not just a Xeon E3 with four cores.

Still, a single quad-core processor today is 40 times more powerful than a 4-way system from twenty years ago (100,000 tpm-C per core is probably possible if TPC-C were still in use, versus 10,000 on a 4-way in 1996). The Xeon E3 could probably support many medium sized organizations. Maximum memory capacity is 64GB (4x16GB unbuffered ECC DIMMs, $130 each). My recollection is that many IO problems went away at the 32-64GB level. And we could still have powerful IO with 2 PCI-E x8 SSDs, or even 2 x4’s.

This is some very interesting research.  Joe has a some gaps in his research (meaning that there’s scope for people to expand upon this), but this is 100% worth the read.

Drew Furgiuele begins his project to build an easy button for backups:

I should also pause for a second and talk about wiring hobby boards like this. Good news first: you won’t electrocute yourself on it. I mean, if you do something really dumb like try to wire it underwater or eat it or something then maybe you could but you shouldn’t ever receive a shock while working with a board like this, even plugged in. The bad news is that even though you won’t damage yourself, you could very well damage the board if you just randomly plug things in. Here’s a hard and fast rule: if you’re not an electronics expert or an electrical engineer, leave it to experts to tell you where and how to wire. I’m not calling myself an expert here, but I have sort of a basic understanding of how to wire these things up. The point I’m attempting at making is: if you want to really learn and understand circuit design, there are lots of great resources of where to get started. And it’s quite a rabbit hole to go down, but it’s well worth your time if you want to learn more.

Read the whole thing.  Over a weekend, with your Pi 3.

Drew Furgiuele has an easy button for SQL Server:

Sounds great, right? I bet some of you already already thinking “Oh man, I can’t wait to run the Linux version SQL Server on this thing!” There’s just one really big catch: the CPUs on Pi boards are ARM-architecture based. Unlike modern processors in our desktops and laptops, these chips are more akin to what you find in mobile phones or other small devices. It also means programs you run or write on your computer are probably 32 or 64 bit and designed for Intel or AMD processors. ARM is a completely different architecture, so we can’t upload something to it and expect it to run. Programs have to be designed for it.

Furthermore a lot of “stock” Pi operating system images are Linux based so it can be difficult to write code that interfaces with .NET or Windows-based services. Not that you can’t; you can certainly write bash scripts that make wget or curl requests.

Based on my experiences at least, nothing with Windows IoT was really that easy…  This is an intro post with a shopping list attached, so I am looking forward to Drew’s making everybody’s lives easier on a budget of $98.

Joe Chang looks at the Samsung 960 PRO SSD:

All the previous PCI-E x4 gen3 NVMe SSDs were rated between 2,000-2,500MB/s in large block read. The 960 Pro is rated for 3,500MB/s read. This is pretty much the maximum possible bandwidth for PCI-E x4 gen3. Each PCI-E gen3 lane is 8Gbit/s, but the realizable bandwidth is less. In earlier generation products, an upper bound of 800MB/s realizable per 8Gbit/s nominal signaling rate was typical.

Presumably there was a reason why every PCI-E x4 was in the 2000-2500MB/s bandwidth. It could be that these were 8-channel controllers and the NAND interface was 333-400MB/s. Even though 8 x 400MB/s = 3,200MB/s, it is expected that excess bandwidth is necessary on the downstream side. The could be other reasons as well, perhaps the DRAM for caching NAND meta-data. Intel had an 18-channel controller, which produced 2,400MB/s in the P750 line, and 2,800MB/s in the P3x00 line.

If you’re looking at a test lab server, this might be a good disk for you.

Glenn Berry has a nice post on current and near-future memory and storage technologies:

Over the past couple of years, we have seen the introduction and increasing usage of NVM Express (NVMe) PCIe SSDs based on existing NAND flash technology. These typically have latencies in the 50-100 microsecond range. They also use the newer, much more efficient NVMe protocol and the PCIe interface, giving much better performance than older SAS/SATA SSDs using the old AHCI protocol.

Currently, Hewlett Packard Enterprise (HPE) is selling 8GB NVDIMM modules for their HPE Proliant DL360 Gen9 servers and HPE Proliant DL380 Gen9 servers. These modules have 8GB of DRAM which is backed by 8GB of flash for $899.00, which is pretty expensive per gigabyte. These two-socket servers have 24 memory slots that each support up to 128GB traditional DDR4 DIMMs. Any slots you use for NVDIMM modules won’t be available for regular memory usage. You can use a maximum of 16 memory slots for NVDIMM usage, which gives you a maximum capacity of 128GB. You must use Intel Xeon E5-2600 v4 series processors in order to get official NVDIMM support. Micron is scheduled to release larger capacity 16GB NVDIMMs in October of 2016.

Read the whole thing.

Erik Darling says get more RAM:

I’m not saying you need a 1:1 relationship between data and memory all the time, but if you’re not caching the stuff users are, you know, using, in an efficient way, you may wanna think about your strategy here.

• Option 1: Buy some more RAM
• Option 2: Buy an all flash array

You’ll still need to blow some development time on tuning queries and indexes, but hardware can usually bridge the gap if things are already critical.

Looking at hardware is a reasonable approach.  The best bet is to satisfy the most pressing need at the margin.  Sometimes that means more (or better) hardware, sometimes it means tuning queries, and sometimes it means application-level changes to retrieve data differently.