I know that forced flushing due to 100% dirty pages causes write IO to halt due to the cache filling, but i've been reading in doco that it actually causes all IO to halt. If this is correct, does that mean read IO is affected by forced flushing as well? If so, why?
>i've been reading in doco that it actually causes all IO to halt.
I'm not aware of any document that says I/O halts. Host response time is likely to increase. However, as cache pages are flushed, which can be very quick, they become available for I/Os.
I recommend you read the 'Write cache management' section of EMC Unified Storage System Fundamentals for Performance and Availability. That section contains a detailed description of the different types of flushing that effect cache, and their effect on performance. The Fundamentals document is available on Powerlink.
Force Flushing begins when the Write Cache% Dirty Pages exceeds the High Watermark.
If your watermarks are set to 60/80 (the normal default), when % Dirty Pages exceeds 80%, the cache will start to force flush.
Forced flushes also create space for new I/Os, though they affect overall performance. When forced flushing takes place, all read and write operations are halted to clear space in the write cache. The time taken for a forced flush is very short (milliseconds), and the array may still deliver acceptable performance.
When the cache dirty pages climbs above the high watermark (80% in this case), “High Watermark Flushing” starts which is a normal flushing process and does not affect host IO.
Forced Flushing occurs when the cache is full (ie: dirty pages at 100%) which means that all pages in write cache differ from what is on disk. Flushing updates the disk to reflect what was written in cache after which the data in cache (no longer dirty) can be overwritten with new incoming writes.
Forced Flushing itself is a high priority process and will impact performance in some way, it could be slight and unnoticeable or high impact. It mostly depends on the load on the system at the time and what the hosts are trying to do.
Read IO can still be serviced during forced flushing, but could be slower due to the high priority forced flushing process consuming the array’s CPU, bus, and disk resources.
Write IO that bypasses cache (due to write aside or write cache disabled status) would still be serviced, again it could be slower, and as long as there are non-dirty pages in write cache, write IO’s will be serviced into cache even during forced flushing. There could be a period of time right at the beginning of the force flush action where there are no free pages and host IO may pend for a few ms. This would translate to host slowness in most cases.
jps00
2 Intern
•
392 Posts
925
0
Posted February 16th, 2012 06:00
>i've been reading in doco that it actually causes all IO to halt.
I'm not aware of any document that says I/O halts. Host response time is likely to increase. However, as cache pages are flushed, which can be very quick, they become available for I/Os.
I recommend you read the 'Write cache management' section of EMC Unified Storage System Fundamentals for Performance and Availability. That section contains a detailed description of the different types of flushing that effect cache, and their effect on performance. The Fundamentals document is available on Powerlink.