Re: Best configuration for bcache/md cache or other cache using ssd

[Stan Hoeppner <stan@hardwarefreak.com>]

On 9/18/2013 12:33 PM, Roberto Spadim wrote:
> Well the internet link here is 100mbps, i think the workload will be a
> bit more than only 100 users, it's a second webserver+database server
> He is trying to use a cheaper server with more disk performace, Brazil
> costs are too high to allow a full ssd system or 15k rpm sas harddisks
> For mariadb server i'm studing if the thread-pool scheduler will be
> used instead of one thread per connection but "it's not my problem"
> the final user will select what is better for database scheduler
> In other words i think the work load will not be a simple web server
> cms/blog, i don't know yet how it will work, it's a black/gray box to
> me, today he have sata enterprise hdd 7200rpm at servers (dell server
> r420 if i'm not wrong) and is studing if a ssd could help, that's my
> 'job' (hobby) in this task

Based on the information provided it sounds like the machine is seek
bound.  The simplest, and best, solution to this problem is simply
installing a [B|F]BWC RAID card w/512KB cache.  Synchronous writes are
acked when committed to RAID cache instead of the platter.  This will
yield ~130,000 burst write TPS before hitting the spindles, or ~130,000
writes in flight.  This is far more performance than you can achieve
with a low end enterprise SSD, for about the same cost.  It's fully
transparent and performance is known and guaranteed, unlike the recent
kernel based block IO caching hacks targeting SSDs as fast read/write
buffers.

You can use the onboard RAID firmware to create RAID1s or a RAID10, or
you can expose each disk individually and use md/RAID while still
benefiting from the write caching, though for only a handful of disks
you're better off using the firmware RAID.  Another advantage is that
you can use parity RAID (controller firmware only) and avoid some of the
RMW penalty, as the read blocks will be in controller cache.  I.e. you
can use three 7.2K disks, get the same capacity as a four disk RAID10,
with equal read performance and nearly the same write performance.

Write heavy DB workloads are a post child for hardware caching RAID devices.

-- 
Stan




> 2013/9/18 Drew <drew.kay@gmail.com>:
>> On Wed, Sep 18, 2013 at 8:51 AM, Roberto Spadim <roberto@spadim.com.br> wrote:
>>> Sorry guys, this time i don't have a full knowledge about the
>>> workload, but from what he told me, he want fast writes with hdd but i
>>> could check if small ssd devices could help
>>> After install linux with raid1 i will install apache mariadb and php
>>> at this machine, in other words it's a database and web server load,
>>> but i don't know what size of app and database will run yet
>>>
>>> Btw, ssd with bcache or dm cache could help hdd (this must be
>>> enterprise level) writes, right?
>>> Any idea what the best method to test what kernel drive could give
>>> superior performace? I'm thinking about install the bcache, and after
>>> make a backup and install dm cache and check what's better, any other
>>> idea?
>>
>> We still need to know what size datasets are going to be used. And
>> also given it's a webserver, how big of a pipe does he have?
>>
>> Given a typical webserver in a colo w/ 10Mbps pipe, I think the
>> suggested config is overkill. For a webserver the 7200 SATA's should
>> be able to deliver enough data to keep apache happy.
>>
>> In the database side, depends on how intensive the workload is. I see
>> a lot of webservers where the 7200's are just fine because the I/O
>> demands from the database are low. Blog/CMS systems like wordpress
>> will be harder on the database but again it depends on how heavy the
>> access is to the server. How many visitors/hour does he expect to
>> serve?
>>
>>
>> --
>> Drew
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-raid" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 
> 
>