Cache: ARC Accesses

Language:

The ARC is the Adaptive Replacement Cache, and is an in-DRAM cache for filesystem and volume data. This statistic shows accesses to the ARC, and allows its usage and performance to be observed.

When to Check ARC Accesses

You can check ARC accesses when investigating performance issues, to understand how well the current workload is caching in the ARC.

ARC Accesses Breakdowns

The available breakdowns of Cache ARC accesses are:

Table 14 Breakdowns of ARC Accesses

Breakdown	Description
hit/miss	The result of the ARC lookup. hit/miss states are described in the table below.
filename	The filename that was requested from the ARC. Using this breakdown allows hierarchy mode to be used, so that filesystem directories can be navigated.
L2ARC eligibility	This is the eligibility of L2ARC caching, as measured at the time of ARC access. A high level of ARC misses which are L2ARC eligible would suggest that the workload would benefit from 2nd level cache devices.
project	This shows the project which is accessing the ARC.
share	This shows the share which is accessing the ARC.
LUN	This shows the LUN which is accessing the ARC.

As described in Execution Performance Impact, breakdown such as by filename would be the most expensive to leave enabled.

The hit/miss states are:

Table 15 Hit/Miss Breakdowns

Hit/Miss Breakdown	Description
data hits	A data block was in the ARC DRAM cache and returned.
data misses	A data block was not in the ARC DRAM cache. It will be read from the L2ARC cache devices (if available and the data is cached on them) or the pool disks.
metadata hits	A metadata block was in the ARC DRAM cache and returned. Metadata includes the on-disk filesystem framework which refers to the data blocks. Other examples are listed below.
metadata misses	A metadata block was not in the ARC DRAM cache. It will be read from the L2ARC cache devices (if available and the data is cached on them) or the pool disks.
prefetched data/metadata hits/misses	ARC accesses triggered by the prefetch mechanism, not directly from an application request. More details on prefetch follow.

Metadata

Examples of metadata:

Filesystem block pointers
Directory information
Data deduplication tables
ZFS uberblock

Prefetch

Prefetch is a mechanism to improve the performance of streaming read workloads. It examines I/O activity to identify sequential reads, and can issue extra reads ahead of time so that the data can be in cache before the application requests it. Prefetch occurs before the ARC by performing accesses to the ARC - bear this in mind when trying to understand prefetch ARC activity. For example, if you see:

Table 16 Prefetch Types

Type	Description
prefetched data misses	Prefetch identified a sequential workload, and requested that the data be cached in the ARC ahead of time by performing ARC accesses for that data. The data was not in the cache already, and so this is a "miss" and the data is read from disk. This is normal, and is how prefetch populates the ARC from disk.
prefeteched data hits	Prefetch identified a sequential workload, and requested that the data be cached in the ARC ahead of time by performing ARC accesses for that data. As it turned out, the data was already in the ARC - so these accesses returned as "hits" (and so the prefetch ARC access wasn't actually needed). This happens if cached data is repeatedly read in a sequential manner.

After data has been prefetched, the application may then request it with its own ARC accesses. Note that the sizes may be different: prefetch may occur with a 128 Kbyte I/O size, while the application may be reading with an 8 Kbyte I/O size. For example, the following doesn't appear directly related:

Data hits: 368
Prefetch data misses: 23

However it may be: if prefetch was requesting with a 128 KByte I/O size, 23 x 128 = 2944 Kbytes. And if the application was requesting with an 8 Kbyte I/O size, 368 x 8 = 2944 Kbytes.

Further Analysis

To investigate ARC misses, check that the ARC has grown to use available DRAM using Cache ARC size.