This chapter provides an overview of the physical and architectural aspects of your Origin 3900 server. System configurations and components are described and illustrated. This chapter includes the following sections:
The SGI Origin 3900 server is the latest model in the SGI Origin 3000 family of servers. It can range from 4 MIPS processors and 1 GB of memory to 512 MIPS processors and 1024 GB of memory. Furthermore, the SGI Origin 3900 servers can be clustered to increase the number of processors from 512 to thousands of processors.
Like all models of the Origin 3000 family of servers, this server is based on SGI NUMAflex shared-memory architecture. The main differences between this server and the other Origin 3000 series servers are as follows:
The Origin 3900 server uses a new packaging scheme that offers four times the processor/memory density over the other Origin 3000 series servers. The Origin 3900 can have up to 128 processors and 256 GB of memory in a single rack; thus, enabling you to have a high-productivity supercomputer in a smaller footprint. For example, a 512-processor Origin 3800 server requires 16 compute racks. An Origin 3900 server that has 512 processors requires only 4 compute racks and 2 router racks (see Figure 2-1).
The Origin 3900 server uses an enhanced system topology (also referred to as a fat-tree topology) that provides two times the bandwidth capabilities over the hypercube topology, which is the topology used by the other Origin 3000 series servers.
The hardware of the Origin 3900 server is fully compatible with the existing Origin 3000 series servers; therefore, you can upgrade an existing Origin 3000 series server with the new Origin 3900 hardware. This upgrade does require converting the existing hypercube topology to the enhanced topology.
The Origin 3900 sever is based on SGI NUMAflex architecture: the third-generation shared-memory system architecture that is the basis of SGI HPC servers and supercomputers. The NUMAflex architecture is specifically engineered to provide technical professionals with superior performance and scalability in a design that is easy to deploy, program, and manage. It has the following features:
Shared access of processors, memory, and I/O -The NUMAflex architecture contains two key components that enable applications to share the processors, memory, and I/O devices: the Bedrock ASIC and the NUMAlink interconnect (see Figure 2-2).
Each Bedrock ASIC in the system is an 8-input by 6-output crossbar that acts as the memory controller between processors and memory in the system for both local and remote memory accesses.
The NUMAlink interconnect channels information between all the bricks in the system to create a single contiguous system memory of up to 1 TB and enables every processor in a system direct access to every I/O slot in the system.
Together, the Bedrock ASICs and the NUMAlink interconnect enable efficient access to processors, local and remote memory, and I/O devices without the bottlenecks associated with switches, backplanes, and other commodity interconnect technologies.
System scalability - The NUMAflex architecture incorporates a low-latency, high-bandwidth interconnect that is designed to maintain performance as it scales in the following dimensions: computing, I/O, and memory. For example, the computing dimension can range from 4 to 512 processors in a single system image (SSI).
Efficient resource management--The NUMAflex architecture is designed to run complex models, and because the entire memory space is shared, large models fit into memory with no programming restrictions. Rather than waiting for all of the processors to complete their assigned tasks, the system dynamically reallocates memory, resulting in faster time to solution.
The SGI Origin 3900 server features the following standard components (see Figure 2-4):
Cx-brick - This compute brick contains 4, 8, 12, or 16 MIPS processors, 1 GB to 32 GB of memory, and a router board. The Cx-brick is equivalent to four C-bricks and one R-brick in one enclosure. For more information about the Cx-brick, see Chapter 3, “Cx-brick”.
IX-brick - This I/O brick is a Crosstalk-to-PCI-X based I/O expansion subsystem that provides the base I/O functionality for the system. For more information about the IX-brick, see Chapter 4, “IX-brick”.
R-brick - This router brick transfers messages between the Cx-bricks via the NUMAlink interconnect. The R-brick is required for systems that contain more than four Cx-bricks. For more information about the R-brick, see Chapter 7, “R-brick”.
L2 controller - This rack-level controller manages the bricks within a rack (not shown in Figure 2-4). For more information about the L2 controller, see Chapter 8, “System Control ”.
Power components - The power components of the Origin 3900 server consist of power distribution units (PDUs) and power bays. The PDUs filter input power, protect against over-current conditions, and remove power from the racks (not shown in Figure 2-4). The power bay inputs AC voltage and outputs 48 VDC and 12 VDC standby voltage to the bricks within a rack. For more information about these power components, see Chapter 9, “Power Components”.
Rack - This industry-standard 19-inch rack contains 39 units (Us) of vertical space that can seat bricks and power bays. (Each U is equivalent to 1.75 inches.) This rack uses two parallel L-shaped mounting rails to support each brick within the rack.
The 39U rack, as shown in Figure 2-4, has front and rear doors that have keylocks to prevent unauthorized access to the system. The rack also has cable entry/exit areas at the bottom of the rack and cable management hardware in the rear of the rack.
The 39U rack is mounted on four casters, two of which are swivel castors. The castors enable the rack to be rolled out of a shipping crate and to its placement at the customer site. The base of the rack has seismic tie-down attachment points and leveling pads. For mechanical specifications of the 39U rack, see Appendix A, “Technical Specifications and Pinouts”.
The Origin 3900 server can contain the following optional components:
System console - This console enables you to power on and power off the system, monitor the system, and enter L1 controller and L2 controller commands.
I/O bricks - The Origin 3900 server supports the following I/O bricks: IX-brick, PX-brick, and X-brick. The IX- and PX-bricks are Crosstalk-to-PCI-X based I/O expansion subsystems that can seat up to 12 PCI-X cards. The difference between these two brick types is that the IX-brick has a DVD-ROM and one or two SCSI disk drives. For more information about the IX-brick, see Chapter 4, “IX-brick”. For more information about the PX-brick, see Chapter 5, “PX–brick”.
The X-brick provides up to four XIO slots for your SGI Origin 3900 server. For more information about the X-brick, see Chapter 6, “X-brick”.
Disk storage components - The Origin 3900 supports the following disk storage components:
For a Fibre Channel solution that supports JBOD (just a bunch of disks) configurations, SGI offers the D-brick2. For more information about the D-brick2, see Chapter 11, “D-brick2 ”.
For a SCSI (small computer system interface) JBOD solution, SGI offers the SGI Total Performance 900 (TP900) storage module. For more information about the TP900, see Chapter 10, “SGI TP900 Storage Module”.
For a Fibre Channel solution that supports both JBOD and RAID configurations, SGI offers the SGI Total Performance 9100 [2Gb TP9100] storage system. For more information about the 2Gb TP9100, see the SGI Total Performance 9100 (2Gb TP9100) Storage System User's Guide (007-4522-00x).
For a Fibre Channel RAID solution, SGI offers the SGI Total Performance 9400 (TP9400) storage system and the SGI Total Performance 9500 (TP9500) storage system. For more information about the TP9400 and TP9500 storage systems, see the SGI Total Performance 9400 and SGI Total Performance 9500 RAID User's Guide (007-4304-00x).
Tape storage components - The Origin 3900 server supports the following tape drives:
DLT 8000 - 8000 native (6-MB/s transfer rate, 40-GB storage capacity)
DLT 8000 - 7000 emulation (5-MB/s transfer rate, 35-GB storage capacity)
LTO (15-MB/s transfer rate, 100-GB storage capacity)
SDLT220 (11-MB/s transfer rate, 110-GB storage capacity)
SDLT320 (16-MB/s transfer rate, 160-GB storage capacity)
T9840 (10-MB/s transfer rate, 20-GB storage capacity)
T9840B (20-MB/s transfer rate, 20-GB storage capacity)
T9940B (30-MB/s transfer rate, 200-GB storage capacity)
For detailed information on the supported tape drives, see the following website: http://sales.corp.sgi.com/products/storage/stk/matrices.html
Graphics components - The Origin 3900 server supports InfinitePerformance or InfiniteReality graphics. The main components of the InfinitePerformance graphics system are the VPro V12 graphics board, the V-brick, and the compositor (CBOB). For more information about the InfinitePerformance graphics components, see Chapter 12, “InfinitePerformance Graphics Components”. The main components of the InfiniteReality graphics system are the G-brick and the N-brick, which are discussed in detail in Chapter 13, “InfiniteReality Graphics Components”.
The Origin 3900 server supports a wide range of system configurations to serve your computing needs (see Table 2-1). The Origin 3900 server scales in 4-processor increments and in two Cx-brick increments. (The Cx-bricks do not have to be fully populated with processors.) The Origin 3900 server also supports a memory-only Cx-brick that enables you to increase the amount of memory without increasing the number of processors.
Table 2-1. Configuration Specifications
System Size | Maximum Number of Processors | Maximum Amount of Memory in GB | Required Number of Metarouters | Required Number of Racks* |
|---|---|---|---|---|
1 Cx-brick | 16 | 32 | 0 | 1 |
2 Cx-bricks | 32 | 64 | 0 | 1 |
4 Cx-bricks | 64 | 128 | 0 | 1 |
6 Cx-bricks | 96 | 192 | 2 | 1 or 2 |
8 Cx-bricks | 128 | 256 | 2 | 2 |
10 Cx-bricks | 160 | 320 | 4 | 2 or 3 |
12 Cx-bricks | 192 | 384 | 4 | 2 or 3 |
14 Cx-bricks | 224 | 448 | 4 | 3 or 4 |
16 Cx-bricks | 256 | 512 | 4 | 3 or 4 |
18 Cx-bricks | 288 | 576 | 18 | 5 or 6 |
20 Cx-bricks | 320 | 640 | 18 | 5 or 6 |
22 Cx-bricks | 352 | 704 | 20 | 5 or 6 |
24 Cx-bricks | 384 | 768 | 20 | 5 or 6 |
26 Cx-bricks | 416 | 832 | 22 | 6 |
28 Cx-bricks | 448 | 896 | 22 | 6 |
30 Cx-bricks | 480 | 960 | 24 | 6 |
32 Cx-bricks | 512 | 1024 | 24 | 6 |
* This number does not include I/O and storage racks. |
|
|
|
|
The following configuration rules apply to the Origin 3900 servers:
When the server contains more than four Cx-bricks, the server must be configured with metarouters (also referred to as R-bricks). For example, servers that contain 6 or 8 Cx-bricks require two metarouters. The third column of Table 2-1 indicates the number of required metarouters for the supported system sizes. The system topology supports a 200 MB/s per processor bisection bandwidth from 32 processors to 512 processors.
Each server or partition within the server must contain one system IX-brick that provides the base I/O functionality.
Each server supports up to eight I/O bricks.
When the server contains a memory-only Cx-brick, the memory-only Cx-brick must be paired with a Cx-brick that has at least four processors.
You can configure your server as a single system, or you can divide it into partitions that operate as separate systems. For example, you can configure an Origin 3900 server that has 512 processors as a single 512-processor system, or you can divide it into as many as eight separate systems. The size of a partition can be configured to match your job size. For more information about partitioning, see Partitioning for Systems that Contain Cx-bricks (007-4426-00x).
An Origin 3900 server is configured so that the server resides in the smallest footprint possible and so that the Cx-bricks are placed low within the racks. The number of compute racks is determined by the number of Cx-bricks, I/O bricks, and storage bricks in the server. For example, the smallest footprint for a 10 Cx-brick system is two compute racks; however, if an additional rack is required to house the I/O and storage bricks, a compute rack, rather than an I/O rack, is added to the system. The Cx-bricks are placed low within the three compute racks and the I/O and storage bricks are placed above the Cx-bricks (see Figure 2-5 and Figure 2-6).
