Hardware Reference Configuration Guide¶
This guide recommends hardware MapD Analytics Platform configurations.
The amount of data you can process with the MapD database is primarily driven by the amount of GPU RAM and CPU RAM available across MapD cluster servers. For absolute zero-latency queries, the system should be able to cache compressed versions of the row- and column-queried fields into GPU RAM, which is called hot data (see Hot Records and Columns). Semi-Hot data utilizes CPU RAM for certain parts of the data.
The Example Systems section lists example configurations to help you make decisions about how to configure your system.
NVIDIA Pascal GPUs¶
Currently, the optimal GPUs on which to run the MapD Analytics Platform are:
- NVIDIA Tesla P40
- Tesla P100
The configurations in this document assume either of these GPUs as your system building block. This does not imply that other GPUs, such as the NVIDIA Tesla K80 or Titan X, will not work, but for production systems Tesla-designated enterprise-grade cards with the latest Pascal architecture, denoted by the “P” prefix, are recommended.
The Tesla P40 and Tesla P100 have advantages in different use cases. For scaling to bigger datasets in a smaller server footprint, the P40 is preferred because of memory size. However, the P100 has significantly higher memory bandwidth and much higher double-precision compute performance, which is ideal for workloads such as the degree of precision needed for geospatial analytics. Also, the Tesla P100 is the only GPU that supports the faster NVLink interconnect, which can provide significant speed increase for some query workloads.
Ask your MapD sales representative for tailored advice on the optimal GPU hardware for your particular use case.
MapD Core Database Architecture¶
Before delving into hardware sizing, it is important to understand at a high level the MapD Core database architecture.
The MapD Core database is a hybrid compute architecture that utilizes GPU, CPU, and storage. GPU and CPU are the Compute Layer, and SSD storage is the Storage Layer.
When thinking about sizing your system, it is important think about the compute and storage layers, separately. Loading raw data into the MapD Core database ingests data onto disk so you can load as much data as you have disk space available, allowing some overhead.
When queries are executed, GPU and CPU RAM become a factor. For a query, the MapD Core database optimizer utilizes GPU RAM first if it is available. You can view GPU RAM as an L1 cache conceptually similar to modern CPU architectures. The MapD Core database attempts to cache the hot data. If GPU RAM is unavailable or filled, the MapD Core database optimizer utilizes CPU RAM (L2). If both L1 and L2 are filled, query records overflow to disk (L3). SSDs are recommended for the Storage Layer to minimize latency.
You can run a query on a record set that spans both GPU RAM and CPU RAM as shown in the above diagram, which also shows the relative performance improvement you can expect based on whether the records all fit into L1, a mix of L1 and L2, only L2, or some combination of L1, L2, and L3.
Hot Records and Columns¶
The Hardware Sizing Schedule table refers to hot records, which are the number of “average-sized” records that you want to put into GPU RAM to get zero-lag performance when querying and interacting with the data. These numbers assume a maximum of 16 hot columns, which is the number of columns involved in the predicate or computed projections (such as, column1 / column2) of any one of your queries. A 15 percent GPU RAM overhead is reserved for rendering buffering and intermediate results. If your queries involve more columns, the number of records you can put in GPU RAM decreases, accordingly.
The server is not limited to any number of hot records. You can store as much data on disk as you desire. The system can also store and query records in CPU RAM, but with higher latency. The hot records represent the number of records on which you can perform zero-latency queries.
The MapD Core database does not require all queried columns to be processed on the GPU. Non-aggregate projection columns, such as
SELECT x, y FROM table, do not need to be processed on the GPU so can be stored in CPU RAM. The Hardware Sizing Schedule CPU RAM sizing assumes that up to 24 columns are used in only non-computed projections, in addition to the Hot Records and Columns.
The amount of CPU RAM needed should be four to eight times the amount of total available GPU memory. Each NVIDIA Tesla P40 has a total of 24GB of on-board RAM available, so if you determine that your application needs four NVIDIA P40 cards, you need between
4 x 24GB x 4 = 384GB and
4 x 24GB x 8 = 768GB of CPU RAM. This correlation between GPU RAM and CPU RAM is because the MapD database uses CPU RAM in certain operations for columns that are not filtered or aggregated.
A MapD Core database deployment should be provisioned with enough SSD storage to reliably store the needed data on disk in compressed format. SSD drives, like the Intel® SSD DC S3610 Series, mat2.5in SATA 6Gb/s, or similar in any size that fits your requirement is recommended. Administrators should size the amount of disk with at least 30 percent overhead because the MapD database uses temporary disk space for various database operations. If you plan to first copy the raw source file(s) to the MapD database server, you need to ensure space is available for both source file(s) and MapD database files.
Hardware Sizing Schedule¶
This schedule estimates the number of records you can process based on GPU RAM and CPU RAM sizes. This applies to the compute layer. For the storage layer, you should provision your application according to SSD Storage guidelines.
|GPU Count||GPU RAM (GB)||CPU RAM (GB)||“Hot” Records|
|(NVIDIA P40)||8x GPU RAM||L1|
If you already have your data in a database, you can look at the largest fact table, get a count of those records, and compare that with the this schedule.
If you have a .csv file, you need to get a count of the number of lines and compare it with this schedule.
MapD uses the CPU in addition to the GPU for some database operations. More cores provide better performance but increase the cost. GPUs are the primary performance driver because the CPU is utilized secondarily. The Intel CPU offerings with 10 cores offer good performance at an affordable price, so you could configure your system with a single NVIDIA P40 GPU and two 10-core CPUs. Similarly, you can configure a server with eight P40s and two 10-core CPUs.
Suggested CPU offerings:
- Intel® Xeon® E5-2650 v3 2.3GHz,10 cores
- Intel® Xeon® E5-2660 v3 2.6GHz, 10 cores
- Intel® Xeon® E5-2687 v3 3.1GHz,10 cores
- Intel® Xeon® E5-2667 v3 3.2GHz, 8 cores
PCI Express (PCIe)¶
GPUs have traditionally been connected to the motherboard using PCIe slots. The PCIe connection is based on the concept of a lane, which is a single-bit, full-duplex, high-speed serial communication channel. The most common numbers of lanes are x4, x8, and x16. The current PCIe 3.0 version with a x16 connection has a bandwidth of 16 GB/s. PCIe 2.0 bandwidth is half the PCIe 3.0 bandwidth, and PCIe 1.0 is half the PCIe 2.0 bandwidth. It is important to use a motherboard that supports the highest bandwidth, preferably, PCIe 3.0. To achieve maximum performance, both the GPU and the PCIe controller should have the same version number.
The PCIe specification permits slots with different physical sizes, depending on the number of lanes connected to the slot. For example, a slot with a x1 connection uses a smaller slot, saving space on the motherboard. However, bigger slots can actually have fewer lanes than their physical designation. For example, motherboards can have x16 slots connected to x8, x4, or even x1 lanes. With bigger slots it is important to know if their physical sizes correspond to the number of lanes. Additionally, some slots may downgrade speeds when lanes are shared. The most common scenario is on motherboards with two or more x16 slots. Some motherboards have only 16 lanes connecting the first two x16 slots to the PCIe controller. This means that when you install a single GPU it has the full x16 bandwidth available, but when two installed GPUs each have x8 bandwidth.
MapD recommends installing GPUs in motherboards with support for as much PCIe bandwidth as possible. On modern Intel chip sets, each socket (CPU) offers 40 lanes, so with the correct motherboards, each GPU can receive x8 of bandwidth. All recommended Example Systems have motherboards designed for maximizing PCIe bandwidth to the GPUs.
MapD does not recommend adding GPUs to a system that is not certified to support the cards. For example, when you want to run eight GPU cards in a machine, the BIOS needs to be aware of the additional address space required for the number of cards. There are also power routing, power supply rating, and air movement through the chassis and cards for thermal control considerations.
The NVLink section describes an emerging alternative technology.
There are a variety of hardware manufacturers today who make suitable GPU offerings, such as Dell, HP, Supermicro, IBM, and NVIDIA. These tables show the relevant configuration items.
2 GPU 2U Server by Dell¶
Description: 256GB RAM, 2-socket CPU, 3.125TB SATA SSD, 2 NVIDIA P40, 2U
|16||R73X||PowerEdge R730 Server|
|16||26404||Intel® Xeon® E5-2640 v4 2.4GHz,25M Cache,8.0GT/s QPI,Turbo,HT,10C/20T (90W) Max Mem 2133MHz|
|16||32GBMM||32GB RDIMM, 2400MT/s, Dual Rank, x4 Data Width|
|16||8006GP||800GB Solid State Drive SATA Write Intensive 6Gbps 2.5in Hot-plug Drive, S3710|
|16||NV24GB||NVIDIA Tesla P40 24GB GPU, Passive|
4 GPU 1U Server by Supermicro¶
Description: 768GB RAM, 2-socket CPU, 6.4TB SATA SSD, 4 NVIDIA P40s, 1U
|2||P4X-E52667V3-SR203||Intel Xeon Processor E52667V3 3.2GHz (Haswell)|
|16||MEM-DR432LHL01-ER21||16GB PC4-19200 2400MHz DDR4 ECC Registered DIMM|
|4||SSDSC2BX016T401||Intel® SSD DC S3610 Series (1.6TB, 2.5in SATA 6Gb/s, 20nm, MLC) 7mm|
|4||AOC-GPU-NVP40M||NVIDIA® Tesla™ P40 GPU Computing Accelerator - 24GB GDDR5 - Passive Cooler|
Workstation NVIDIA DGX Station¶
Description: 256GB RAM, 1 CPU, 8TB SATA SSD, 4 NVIDIA V100s NVLINK
|1||Intel Xeon Processor E52698V4 2.2GHz (20 corel)|
|256GB LRDIMM DDR4|
Data: 3X 1.92 TB SSD RAID 0
OS: 1X 1.92 TB SSD
8 GPU 4U Server by Supermicro¶
Description: 768GB RAM, 2-socket CPU, 6.4TB SATA SSD, 8 NVIDIA P40s, 4U
|2||Eight-Core Intel® Xeon® Processor E5-2667 v4 3.20GHz 25MB Cache|
|24||32GB PC4-19200 2400MHz DDR4 ECC Registered Load-Reduced DIMM|
|4||1.6TB Intel® SSD DC S3520 Series 2.5” SATA 6.0Gb/s Solid State Drive|
|8||NVIDIA® Tesla™ P40 GPU Computing Accelerator - 24GB GDDR5 - Passive Cooler|
2 GPU Developer Workstation from System76¶
Description: 64GB RAM, 1-socket CPU, 1TB SATA SSD, 2 NVIDIA 1080Ti
|4.0 GHz i7-6850K (3.6 up to 4.0 GHz – 15 MB Cache – 6 Cores – 12 threads)|
|64 GB Quad Channel DDR4 at 2400MHz (4× 16 GB)|
|1 TB 2.5” SSD|
|2||11 GB GTX 1080 Ti with 3584 CUDA Cores|
9 GPU 4U Server by HP¶
Description: 1TB RAM, 4-socket CPU, 7.68TB SATA SSD, 9 NVIDIA P40s, 4U
|1||HPE DL580 Gen 9||HP ProLiant DL580 G9 4U Rack Server|
|4||E7-8890v3||Intel® Xeon® E7-8890 v3 (18 core, 2.5 GHz, 45MB, 165W)|
|16||726724-B21||HP 64GB (1x64GB) Quad Rank x4 DDR4-2133 CAS-15-15-15 Load Reduced Memory Kit|
|2||816929-B21||3.84TB 6G SATA Read Intensive-3 SFF 2.5-in SC Solid State Drive|
|9||Q0V80A||NVIDIA® Tesla™ P40 GPU|
8 GPU 4U Server by NVIDIA¶
Description: 512GB RAM, 2-socket CPU, 7.68TB SATA SSD, 8 NVIDIA V100 NVLINK, 4U
|2||20-Core Intel Xeon E5-2698 v4 2.2 GHz|
|512 GB 2,133 MHz DDR4 LRDIMM|
|4||1.92 TB SSD RAID 0|
GPU 2U IBM Power System S822LC Server by IBM¶
Description: 512GB RAM, 2-socket CPU, 2TB SATA SSD, 4 NVIDIA P100 NVLINK, 2U
|1||2x 8 core CPUs at 3.25 GHz (16x POWER8 cores with NVIDIA NVLink)|
|2||10 cores (2 x 10c chips) / 160 threads, POWER8 with NVLink|
|2||1.92TB 2.5” 6GB/s SATA 512N FORMAT SSD|
|4||NVIDIA Tesla P100 SXM2 GPUs with NVIDIA NVLink|
Penguin Computing: http://www.penguincomputing.com/products/gpu-servers/