DIY ZeusRAM drive

[Joe Reid]

I had some time to kill this afternoon while I was waiting for some users and their management to respond to some questions about their new feature and I came across this article:

https://forums.servethehome.com/index.php?threads/poor-mans-diy-zeusram.2712/

The basic idea is to take a battery backed write cache RAID controller and gave it a single ssd volume (partitioned to the size of the write cache?) and put SLOG/ZIL on that.

After I finished reading it I looked around and found this:
http://www.newegg.com/Product/Product.aspx?Item=9SIA6ZP3K25609

Any comments on his idea? Most of my use is home CIFS services with 1 or 2 users. But network write with lots of small files have issues and I'm beginning to look for solutions.

--
joe

[Joe Reid]

On Thursday, January 21, 2016 at 7:11:30 PM UTC-6, Joe Reid wrote:

> The basic idea is to take a battery backed write cache RAID controller and gave it a single ssd volume (partitioned to the size of the write cache?) and put SLOG/ZIL on that.

There is obviously something that is not coming across in drive specifications. Everywhere I read about people looking for a less expensive version of what the ZeusRam "drive" does make comments about the Intel DC S3500/S3700 series of drives. When I look at those specs they don't seem very astounding, compared to the multitude of SSDs out there like say the Mushkin Eco2 60gb. The specs for the Mushkin _seem_ to leave even the S3700 in the dust.

The one spec that Intel provides that no one else does is read and write latency. Intel lists 50 and 65 microseconds, the only thing that resembles that spec is Muskin saying access time is sub 1ms...which is a TON more than 50 or 65us. But the write speeds and 4k random writes speeds seem much lower than Mushkin.

can anyone shed any light on how I should read the specs to understand what's going on here? The 100gb Intel drive runs about what the HP400 controller, a 8484 cable and a cheap ssd cost anyway, but it seems like the controller idea would have better longevity as I could keep moving the 512mb partition the controller sees every few months or so...lots of 512mb partitions in a 60gb drive.

--
joe

[Ian Collins]

Please wrap your lines!

Joe Reid wrote:
> On Thursday, January 21, 2016 at 7:11:30 PM UTC-6, Joe Reid wrote:
>
>> The basic idea is to take a battery backed write cache RAID
>> controller and gave it a single ssd volume (partitioned to the size
>> of the write cache?) and put SLOG/ZIL on that.
>
> There is obviously something that is not coming across in drive
> specifications. Everywhere I read about people looking for a less
> expensive version of what the ZeusRam "drive" does make comments
> about the Intel DC S3500/S3700 series of drives. When I look at
> those specs they don't seem very astounding, compared to the
> multitude of SSDs out there like say the Mushkin Eco2 60gb. The
> specs for the Mushkin _seem_ to leave even the S3700 in the dust.

The S3700 is a popular choice for ZFS logs because it combines high 4K
write IOPs with longevity and power fail protection. Few if any
"consumer" SSDs provide power fail protection which is particularly
important for a log device. Its design also provides very consistent
performance numbers: the throughput does not drop away during an
extended test cycle.

> The one spec that Intel provides that no one else does is read and
> write latency. Intel lists 50 and 65 microseconds, the only thing
> that resembles that spec is Muskin saying access time is sub
> 1ms...which is a TON more than 50 or 65us. But the write speeds and
> 4k random writes speeds seem much lower than Mushkin.

Latency is a real performance killer in write intensive use cases. Also
be aware that most SSD numbers (including those for the Eco2) are for
high queue depths not found in the log device use case.

--
Ian Collins

[Joe Reid]

On Friday, January 22, 2016 at 2:07:27 PM UTC-6, Ian Collins wrote:
> Please wrap your lines!

can you log that with the google groups site developers :-)

> The S3700 is a popular choice for ZFS logs because it combines high 4K
> write IOPs with longevity and power fail protection. Few if any
> "consumer" SSDs provide power fail protection which is particularly
> important for a log device. Its design also provides very consistent
> performance numbers: the throughput does not drop away during an
> extended test cycle.

This is where I'm confused. The newegg page for the S3700 says 19k 4k
write IOPS, but the newegg page for the Eco2 60gb says 85k IOPS. So
there's something else in those numbers (or a typo?) I know these
Mushkin drives aren't screamers...

> Latency is a real performance killer in write intensive use cases. Also
> be aware that most SSD numbers (including those for the Eco2) are for
> high queue depths not found in the log device use case.

That's why all those consumer drives say things like "performance numbers
achieved with queue depth of 10"?

Thanks for the info, this is fascinating stuff - reading through the
marketing is the best part of our jobs/hobbies right?

--
joe

[Ian Collins]

Joe Reid wrote:
> On Friday, January 22, 2016 at 2:07:27 PM UTC-6, Ian Collins wrote:
>> Please wrap your lines!
>
> can you log that with the google groups site developers :-)
>
>> The S3700 is a popular choice for ZFS logs because it combines high 4K
>> write IOPs with longevity and power fail protection. Few if any
>> "consumer" SSDs provide power fail protection which is particularly
>> important for a log device. Its design also provides very consistent
>> performance numbers: the throughput does not drop away during an
>> extended test cycle.
>
> This is where I'm confused. The newegg page for the S3700 says 19k 4k
> write IOPS, but the newegg page for the Eco2 60gb says 85k IOPS. So
> there's something else in those numbers (or a typo?) I know these
> Mushkin drives aren't screamers...

I guess it depends on how they measure. If I'm going to include parts
in a system I'm building for a client, I'll get hold of a sample and do
my own testing in a real world system. I wouldn't use any drive without
power fail protection for a log, so those Eco2 wouldn't get a look in.

There's a good detailed review of the 3700 here:

http://www.anandtech.com/show/6433/intel-ssd-dc-s3700-200gb-review/10

The new 3710 drives are better still and cost less...

>> Latency is a real performance killer in write intensive use cases. Also
>> be aware that most SSD numbers (including those for the Eco2) are for
>> high queue depths not found in the log device use case.
>
> That's why all those consumer drives say things like "performance numbers
> achieved with queue depth of 10"?
>
> Thanks for the info, this is fascinating stuff - reading through the
> marketing is the best part of our jobs/hobbies right?

It can consume an awful lot of time :)

--
Ian Collins

[Joe Reid]

On Friday, January 22, 2016 at 4:32:57 PM UTC-6, Ian Collins wrote:
> > Thanks for the info, this is fascinating stuff - reading through the
> > marketing is the best part of our jobs/hobbies right?
>
> It can consume an awful lot of time :)

...please don't tell my boss what I was doing, or not doing, Thrusday afternoon...

--
joe