SPARC System Specific Parameters

The following parameters apply to sun4v and SPARC M-Series sun4u platforms.

consistent_coloring

Description

The ability to use different page placement policies on the UltraSPARC platform is available. A page placement policy attempts to allocate physical page addresses to maximize the use of the L2 cache. Whatever algorithm is chosen as the default algorithm, that algorithm can potentially provide less optimal results than another algorithm for a particular application set. This parameter changes the placement algorithm selected for all processes on the system.

Based on the size of the L2 cache, memory is divided into bins. The page placement code allocates a page from a bin when a page fault first occurs on an unmapped page. The page chosen depends on which of the three possible algorithms are used:

• Page coloring – Various bits of the virtual address are used to determine the bin from which the page is selected. consistent_coloring is set to zero to use this algorithm. No per-process history exists for this algorithm.

• Virtual addr=physical address – Consecutive pages in the program selects pages from consecutive bins. consistent_coloring is set to 1 to use this algorithm. No per-process history exists for this algorithm.

• Bin-hopping – Consecutive pages in the program generally allocate pages from every other bin, but the algorithm occasionally skips more bins. consistent_coloring is set to 2 to use this algorithm. Each process starts at a randomly selected bin, and a per-process memory of the last bin allocated is kept.

Dynamic?

Yes

Validation

None. Values larger than 2 cause a number of WARNING: AS_2_BIN: bad consistent coloring value messages to appear on the console. The system hangs immediately thereafter. A power-cycle is required to recover.

When to Change

When the primary workload of the system is a set of long-running high-performance computing (HPC) applications. Changing this value might provide better performance. File servers, database servers, and systems with a number of active processes (for example, compile or time sharing servers) do not benefit from changes.

Commitment Level

Unstable

tsb_alloc_hiwater_factor

Description

Initializes tsb_alloc_hiwater to impose an upper limit on the amount of physical memory that can be allocated for translation storage buffers (TSBs) as follows:

tsb_alloc_hiwater = physical memory (bytes) / tsb_alloc_hiwater_factor

When the memory that is allocated to TSBs is equal to the value of tsb_alloc_hiwater, the TSB memory allocation algorithm attempts to reclaim TSB memory as pages are unmapped.

Exercise caution when using this factor to increase the value of tsb_alloc_hiwater. To prevent system hangs, the resulting high water value must be considerably lower than the value of swapfs_minfree and segspt_minfree.

Data Type

Integer

Default

32

Range

1 to MAXINIT

Note that a factor of 1 makes all physical memory available for allocation to TSBs, which could cause the system to hang. A factor that is too high will not leave memory available for allocation to TSBs, decreasing system performance.

Dynamic?

Yes

Validation

None

When to Change

Change the value of this parameter if the system has many processes that attach to very large shared memory segments. Under most circumstances, tuning of this variable is not necessary.

Commitment Level

Unstable

default_tsb_size

Description

Selects size of the initial translation storage buffers (TSBs) allocated to all processes.

Data Type

Integer

Default

Default is 0 (8 KB), which corresponds to 512 entries

Range

Possible values are:

Value Description 0 8 KB 1 16 KB 3 32 KB 4 128 KB 5 256 KB 6 512 KB 7 1 Mbyte

Dynamic?

Yes

Validation

None

When to Change

Generally, you do not need to change this value. However, doing so might provide some advantages if the majority of processes on the system have a larger than average working set, or if resident set size (RSS) sizing is disabled.

Commitment Level

Unstable

enable_tsb_rss_sizing

Description

Enables a resident set size (RSS) based TSB sizing heuristic.

Data Type

Boolean

Default

1 (TSBs can be resized)

Range

0 (TSBs remain at tsb_default_size) or 1 (TSBs can be resized)

If set to 0, then tsb_rss_factor is ignored.

Dynamic?

Yes

Validation

Yes

When to Change

Can be set to 0 to prevent growth of the TSBs. Under most circumstances, this parameter should be left at the default setting.

Commitment Level

Unstable

tsb_rss_factor

Description

Controls the RSS to TSB span ratio of the RSS sizing heuristic. This factor divided by 512 yields the percentage of the TSB span which must be resident in memory before the TSB is considered as a candidate for resizing.

Data Type

Integer

Default

384, resulting in a value of 75%. Thus, when the TSB is 3/4 full, its size will be increased. Note that some virtual addresses typically map to the same slot in the TSB. Therefore, conflicts can occur before the TSB is at 100% full.

Range

0 to 512

Dynamic?

Yes

Validation

None

When to Change

If the system is experiencing an excessive number of traps due to TSB misses, for example, due to virtual address conflicts in the TSB, you might consider decreasing this value toward 0.

For example, changing tsb_rss_factor to 256 (effectively, 50%) instead of 384 (effectively, 75%) might help eliminate virtual address conflicts in the TSB in some cases, but will use more kernel memory, particularly on a heavily loaded system.

TSB activity can be monitored with the trapstat -T command.

Commitment Level

Unstable