TAMU Supercomputing Facility -- SGI Altix 3700 Hardware

cosmos.tamu.edu - SGI Altix 3700

cosmos.tamu.edu is an SGI Altix 3700 supercomputer, which consists of 128 Itanium-2® processors, and 256 Gigabytes of main memory. It is based on a Distributed Shared Memory (DSM) architecture, where memory is physically distributed among 32 Computation-bricks (``C-bricks''). Each C-brick consists of two pairs of 1.3GHz Itanium-2 64-bit μ-processors, locally attached memory, cache-coherence logic and interconnection fabric. Altix 3000 class of computers provide a global, cache-coherent, Non-Uniform Access Time (cc-NUMA) shared memory.

Each processor can access any memory locations through a high-bandwidth/low-latency system interconnect, based on SGI's NUMAflex architecture. The interconnect is organized in a Dual Plane, Fat-Tree layout and scales with system growth. Initial memory access times vary with the number of router-hops (distance) that data items have to travel.

cosmos is directly attached to a 10 Terabyte TP9500 RAID system via four 2-Gigabit/sec fibre-channels.

The operating system of cosmos is the IA-64 version of Linux with several SMP and NUMA-aware enhancements provided by SGI and the OSS community. The operating system runs as a Single System Image (SSI) on the 128 PEs of cosmos. We are currently using SGI ProPack 4 which is based on the 2.6 distribution of the Linux kernel.

The information contained in this page has been compiled from SGI manuals and papers published in the literature. See also the architecture part of the cosmos user guide.

Table 1. Altix 3700 Configuration -- Summary

Component Specifications

Number of processors 128

Physical memory size 256 GBytes DDR SDRAM with ECC

Memory architecture Cache Coherent, Non-Uniform Memory Access (cc-NUMA) times

Operating system ProPack 4, based on IA-64 Linux 2.6
128 processors single operating system image (SSI) size

Processor type, ISA 1.3 GHz Intel Itanium-2® μ-processors (Madison), IA-64 Little Endian

Cache memories (on processor die)

Level	Total size	Block Size	Associativity	Write policy	Load/Store latency (Clocks)
L1	16KB instr	64 bytes	4	N/A	1/
L1	16KB data	64 bytes	4	write-through	1/3
L2	256KB unified	128 bytes	8	write-back	5/7
L3	3MB unified	128 bytes	12	write-back	14/7

Number of processors / node, FSB 2 [Fig. 1]

Size of local memory / node 4 Gigabytes [Fig. 1]

Number of nodes and SHUBs / C-brick 2 (NODE 0, NODE 1) [Fig. 2]

Number of C-bricks 32

Cache coherence protocol within node Itanium-2 snoopy bus (FSB)

Interconnection between nodes for global shared memory SGI® NUMAlink™-3
Dual-plane, fat-tree topology [Fig. 3]
Router-brick (R-brick) with 8×8 cross-bar switch [Fig. 4]
3.2 GB/sec bidirectional bandwidth per link
400 MB/sec bandwitdh available for each processor
path length is O(log N), N = number of processors; 5 routing hops maximum; 50 nanosecs / router

Global cache coherence protocol SGI® NUMAflex™ Directory-based, write-invalidation (cache controllers in SHUBs)

Memory-only modules Not on cosmos

Number of IX-bricks 2 (1 for each 64 processors)

Base PCI/PCI-X slots Per IX-brick
5 PCI-X buses (64-bit, 133 MHz), 10 slots available
1 64-bit, 66 MHz PCI slot

Networking Six 1-Gigabit ethernet (2 fiber, 4 copper)

System disks Four 36GB disks (1 for each IX-brick)

Disk expansion unit TP9500 10 Terabyte RAID disk

Fig. 1 An Altix node with 2 Itanium-2 μ-processors on the Front-Side Bus (FSB), local DDR SDRAM, SHUB chip and NUMAlink-3 and XIO channels.

Fig. 2 An Altix C-brick with 2 nodes, 2 NUMAlink-3 and 2 XIO channels.

Fig. 3 An 128 PE Altix with a Dual-Plane Fat-Tree interconnection topology.

Fig. 4 Detail of the internals of a Router-Brick (R-brick) with an 8×8 cross-bar switch and 8 bidirectional NUMAlink channels.

cosmos.tamu.edu - SGI Altix 3700

`cosmos.tamu.edu` - SGI Altix 3700