Identifying Memory Performance Issues (CLI)
Use the following procedure to identify and remedy memory hardware bottlenecks on the
appliance. Based on the results of the analytic dataset, suggested corrective actions are
provided to increase memory performance by installing more DRAM.
-
Create a worksheet as described in Creating a Worksheet (CLI), select that worksheet, and then enter dataset.
hostname:analytics worksheets> select worksheet-000
hostname:analytics worksheet-000> dataset
-
Enter set name=arc.accesses[hit/miss], and then enter
commit to add cache ARC accesses broken down by hit/miss to your
worksheet.
hostname:analytics worksheet-000 dataset (uncommitted)> set name=arc.accesses[hit/miss]
name = arc.accesses[hit/miss]
hostname:analytics worksheet-000 dataset (uncommitted)> commit
-
Enter done, and then enter done again to exit
the context.
hostname:analytics worksheet-000> done
hostname:analytics worksheets> done
-
Wait at least 10 minutes, and then go to analytics datasets.
Note -
Ten minutes is a general guideline. The amount of time may be adjusted if you have
shorter-duration workloads that are memory intensive.
hostname:> analytics datasets
-
Enter show to view a list of available datasets.
hostname:analytics datasets> show
Datasets:
DATASET STATE INCORE ONDISK NAME
dataset-000 active 1.27M 15.5M arc.accesses[hit/miss]
dataset-001 active 517K 9.21M arc.accesses[hit/miss=metadata hits][L2ARC eligibility]
...
hostname:analytics datasets>
-
Enter select and the dataset with the name
arc.accesses[hit/miss].
In this example, dataset name arc.accesses[hit/miss] corresponds to
dataset-000.
hostname:analytics datasets> select dataset-000
-
Enter read 600 to read the last 600 seconds, or 10 minutes, of the
dataset.
hostname:analytics dataset-000> read 600
-
Examine the data.
You may want to install more DRAM when all of the conditions in the following table
are present.
|
|
ARC access hits for data or metadata are at least 75-97% compared to
misses
|
The ARC is providing a benefit by storing data or metadata that the
applications need.
|
ARC access hits for data or metadata are significantly greater than prefetch
hits
|
The majority of the ARC accesses are for real applications rather than just
the prefetch mechanism.
|
ARC is accessed at least 10,000 times per second
|
The appliance is hitting DRAM, which is not the typical utilization of an
idle system.
|
Nearly all memory is consumed by the ARC, leaving very little unused
memory
|
The appliance is utilizing all of the DRAM possible for the ARC, not just
serving a hot workload out of a small subset of the DRAM that is already
present.
|
|