sar -b

[Malcolm McKenzie]

The output from my sar -b command is attached.

I am concerned that the %rcache and %wcache levels are too low.

I'd like an explanation of why they may be low and any steps I can take to improve them.

[Vincent Stedema · Accepted solution]

According to HP, the %rcache should nearly always be 100 and %wcache should never be less than 70. I'll try and find the document which elaborates on these specifications. Anyway, if your rates tend to fall below these thresholds, you might want to increase dbc_max_pct to increase the size of the dynamic buffer cache. If dbc_max_pct is already at 50% and the dynamic buffer cache is completely used (use the ables option in glance), then the machine migth be short on memory.

Vincent

[Thierry Poels_1]

Hi,
I think the the "rabbit"-answer has already been given by Vincent.
But just want to add: if you are mainly using a database like Oracle which does its own buffering then the OS filesystem buffering is less important.
regards,
Thierry.

[Wodisch]

Hello Malcolm,

in case the processes on your box are writing "new"
data to disk (like log-files), then everything is fine -
appending data to existing files cannot be found in the
buffer cache. Your values seem to be really high, not
low, btw!
Be careful NOT to use more than about 400MB of cache,
or it will actually slow down your system.
Use "sar -d" and "iostat" to check for fragmented data
files instead.

HTH,
Wodisch

[Philip Chan_1]

Malcolm,

Here is my 2 cents.

Any read cache hit above 95% should be fine, of course higher the better. Also, don't just look at the percentage alone, the number of physical read/write per second should be taken into account too. As I can see from your sar output that while your %rcache figures dropped below 70%, the number of physical read per sec (pread) was only 2, this I don't think will impose much performance impact on your system.

Usually, putting in more memory will improve system performance because cache hit rate gets better, therefore less physical I/O which is the major factor for a system to run smoothly. This is especially true if you're running RDMBS such as Sybase or Oracle. Large amount of cache memory may be causing more OS management overhead but I believe the benefits on reduced physical I/O could easily outweight such side effect.

Rgds,
Philip

[Philip Chan_1]

Malcolm,

I made a mistake, the physical disk read/write should be represented by bread/s and bwrite/s (not pread/s and pwrite/s as I thought before), One explanation on this is that your system is really lack on memory. Other possibilities are that your system ran RDBMS like Sybase or Oracle, and there were not enough memory allocated. Also, it is possible that there were non-indexed SQL queries which caused table scans, hence led to large amount of physical I/Os.

Rgds,
Philip