Tuesday, September 27, 2011

The SPARC T4

Besides the new core and cache layout the SPARC T4 is very similar to the SPARC T3 in I/O and memory design. It also supports 16 DDR3 DIMMs with two controllers, built on 40nm technology, uses 6 * 9.6GB/s coherency links, Dual PCIe Gen2 interface and dual on-chip 10GbE interfaces.

The new S3 core is whats sets the SPARC T4 apart from previous T-processors. It has several features that greatly improves the performance compared to the T2/T3:
  • Out-oforder execution
  • Dual instruction issue
  • Data/instruction prefetch
  • Deeper pipeline
  • MMU Page size up to 2GB
  • Level 3 cache
"All of these characteristics in the SPARC T4 have yielded improvements in single-thread performance by 5X while retaining networking and throughput performance equal to that of previous multicore processors from Sun/Oracle."

The crypto graphic units (SPU, previously MAU) have also been moved into the pipeline so there is no longer any need to managed crypto units be individually. They provide high performance crypto acceleration for the supported algorithms:

"The SPU is designed to achieve wire-speed encryption and decryption on the processor’s 10 GbE ports. "
"These accelerators support 16 industry standard security ciphers and enable high speed encryption at rates 3 to 5 times that of competing processors."

The T4 has the ability to execute critical threads exclusively on a core. This is done by issuing a system call but it can also be handled from the command line by raising the thread priority to above 60. This means that existing applications can take advantage of this feature without rewrite. Applications that depend on a single high performance thread this thread can be declared as critical while other threads can take advantage of the highly threaded design of the T4 allowing great throughput while still providing the needed single-thread performance.

(I think what they should have been more specific in the first sentence by writing Oracle Solaris 10 and 11)
"Oracle 10 now and 11 (initial release) will have the ability to permit either a user or programmer to allow the Oracle Solaris Scheduler to recognize a 'critical thread' by means of raising its priority to 60 or above through the use of either the Command Line Interface or system calls to a function. If this is done, that thread will run by itself on a single core, garnering all resources of that core for itself. The one condition that would prevent this single thread from executing on a single core is when there are more runnable threads than available CPUs. This limit was put into place to prevent resource starvation to other threads. There will be further enhancements to Critical Thread Optimization done for Oracle Solaris 11 initial release)."

For many organization the high single thread performance of the T4 will enable the T-series to be used as a general platform for SPARC virtualization. Previously you could mix zones or dynamic domains on the M-series with LDOM on the T-series but there was not good solution for general workloads due to the weaker performance of the T2/T3 cores. Solaris zones are still useful tool for virtualization that has it advantages but the built-in virtualization in the T-series can provide better separation and live migration between hosts (which in turn can contain zones).

Besides the new processor the SPARC T4 systems comes pretty much the same chassis as the T3, they do however support the double amount of memory using 16GB DIMMs.

No comments: