[dm-devel] dm-cache: blocks don't get cached on 3.18.21-17.el6.x86_64

[Thanos Makatos]

On 14 March 2016 at 12:44, Joe Thornber <thornber at redhat.com> wrote:
> On Mon, Mar 14, 2016 at 09:54:06AM +0000, Thanos Makatos wrote:
>> (I've already reported this issue to centos and centos-devel. and
>> waited long enough but didn't get any reply.)
>>
>> I'm evaluating dm-cache on CentOS 6 kernels  3.18.21-17.el6.x86_64 (Xen 4) and
>> 2.6.32-573.7.1.el6.x86_64 (KVM). The test I do is a simple sequential
>> read using dd(1). read_promote_adjustment and sequential_threshold
>> have been set to 1 and 0, respectively. For the 2.6.32 kernel, all
>> seems to be working fine, "#usedcacheblocks" correctly reflects the
>> number of cached blocks based on what I'm reading with dd(1),
>> performance is pretty much native SSD performance. However, the same
>> test on the 3.18.21-17.el6.x86_64 kernel results in "#usedcacheblocks"
>> being stuck to "2" and no performance improvement is observed.
>>
>> Any ideas what could be wrong? How can I further debug this?
>
> There may not be anything wrong, dm-cache is very cagey these days
> about promoting blocks to the ssd without evidence that they're
> hotspots.  Hitting blocks once with a dd is not enough.

I've been hitting the same block many many times, still it doesn't get
promoted. Is there a foolproof way that results in blocks getting
cached?