These Servers are TOO EXPENSIVE - Hybrid Storage Explored

Linus Tech Tips ·Linus Tech Tips ·2019-05-06 · 2,223 words · ~11 min read
Floatplane YouTube

Transcript

JSON SRT VTT 187
0:00 so i've got a problem when it comes to building a storage server there are two
0:03 main options for your drives you can either go with ssds which are compact
0:08 and have amazing performance or you can go with traditional mechanical storage
0:14 which has the benefit of lower prices and outstanding capacity with the
0:19 drawback of being an order of magnitude
0:22 or more slower especially in highly
0:26 random workloads like if they were loaded up with a bunch of virtual
0:29 machines or in our case if they're pulling network storage duty for a whole
0:35 bunch of heavy users who are all trying to edit 8k videos at the same time if
0:41 only there were some way to
0:44 merge the benefits of both
0:47 with the drawbacks of neither
0:50 well there is it's called the apple fusion drive
0:54 and this episode is brought to you by ifixit the marlin screwdriver set from
0:58 ifixit features five specialty precision screwdrivers check it out at ifixit.com
1:03 forward slash Linus at the link below
1:13 the concept of tiered storage is nothing new it's been commonplace in the data
1:18 center for many years and has even made its way down to the consumer level in
1:22 the form of Intel's rapid storage technology or rst which uses a small
1:28 fast SSD to accelerate a larger slower
1:32 mechanical drive and more recently Intel's octane memory technology which
1:37 does pretty much the exact same thing but with even faster optane rather than
1:43 nand ssds the problem that i'm running
1:46 into here is that most of the tiering solutions out there have one of two
1:52 problems they are either inordinately complicated to set up requiring
1:57 extensive knowledge of not just Linux but even broader storage architecture or
2:03 they are expensive because you are paying the folks who have the
2:07 aforementioned knowledge to create a turnkey solution for you
2:12 there are some exceptions to this however and i would like to explore a
2:16 couple of them so the popular freebsd based freenas software from ix systems
2:22 uses your system's RAM and its ssds to cache frequently used
2:28 data that's stored on your mechanical hard drives but the issue is that zfs's
2:33 particular implementation is geared more towards accelerating performance of
2:39 large databases and doesn't look like it's going to be great for our workload
2:44 because the file sizes that we want cached that we're storing on our hard
2:48 drives are so large like these are 8k
2:51 red video files and there are practical
2:55 limits to how much level 2 arc that
2:59 would be the SSD tier i mean there's also practical limits to how much RAM
3:03 you can put in a given system
3:06 so then the one we're gonna try today is the
3:10 built-in tiering mode for storage spaces
3:13 in Windows server and what better way to test this out than to set up a tiered
3:19 storage space on the network so i'm going to use this test server right here
3:24 and then have our team of editors just switch over to it and see what happens
3:32 and it's at this point then that we wondered whether or not optane would
3:36 make a difference for some of the heavy tasks that we deal with around the
3:39 office no no it does not now our regular editing
3:44 server uses 31 1.2 terabyte NVMe drives
3:49 and it was really expensive for us to build and they've been complaining that
3:53 there's not enough space on it because we've only got about 24 terabytes of
3:58 usable space after all the lost space to
4:01 redundancy on it and they were like oh well you could just upgrade it i went yeah okay
4:06 technically i do have another 17 bays that i could
4:11 fill up but that's going to cost me like seven grand eight grand even if i buy
4:16 used drives on ebay so i was
4:20 looking for another way so our test server here the config is
4:25 pretty straightforward i've got four 10
4:28 terabyte drives then because i want to see exactly how much
4:34 cash you need this is what really got me thinking about it
4:37 Intel uses just 32 gigs of cash
4:42 for large multi-terabyte hard drives and we've
4:46 seen real world huge increases in performance so i went okay let's try
4:52 this so i took two of their 900 series
4:56 optane pci express drives and then i've got a total of 40 terabytes of
5:01 mechanical storage and i want to see if for a fraction of the price
5:06 we can get good enough performance that
5:09 the editors aren't gonna run into a bottleneck
5:12 so you can see i've got my four mechanical drives i've got my two ssds
5:18 uh allocation is all gonna be automatic so i go ahead and i try to create a pool
5:23 okay i've created a pool now i try to
5:27 create a virtual disk create storage tiers on this virtual disk it should
5:32 automatically detect which ones are your fast drives and which ones are your
5:38 capacity drives so we're going to use mirroring because
5:42 that way we're protected from a drive failure we're going to use fixed
5:46 provisioning and now we get to specify the size of our faster and our standard
5:51 tier we're going to go ahead and select the maximum sizes failed to create
5:55 virtual disks not supported so the issue is that the gui version of storage
5:59 spaces doesn't have some very important
6:02 options like selecting how many columns you want
6:06 so when you only have two drives and the
6:09 default is three columns it spits out this stupid error how did
6:14 nobody at microsoft notice this for
6:17 years anyway doesn't matter
6:20 i called in a lifeline led eater over on the forum as a guru when it comes to
6:25 this stuff so i'm just going to pull up his dm over there and follow his
6:29 instructions to hopefully do this in powershell so this will be fun you guys
6:33 are going to be learning right along with me i haven't done this before the
6:36 video so i can still use the gui to create my storage pool apparently
6:41 i'm going to go ahead and do that then i need to fire up powershell and
6:45 change the default number of columns
6:49 okay well i don't know what i just did
6:52 but i'm going to change it to mirror
6:56 number of columns default one i think i think we did that let's
7:02 find out so i think the command i just entered
7:08 changed the mirror layout default
7:12 to one column it did not
7:17 okay uh so i'm using powershell ise now
7:22 and this seems to be working better
7:26 mirror number of columns default one
7:30 come on so mad right now
7:35 i just i just want to change a stupid
7:39 setting it's definitely called
7:43 mirror it's definitely called storage pool one like i create
7:48 these commands using their own thing
7:52 we'll figure this out at some point the good news is there's a backup plan
7:56 i swear i had this working at one point
8:00 what is this oh okay oops actually yeah
8:05 oops all right
8:09 here we go so right now we've only got a 10 gig link i can put a faster card in
8:14 here if it ends up being a networking bottleneck but for now what we need to
8:18 do is grab a bunch of
8:22 just like video projects that are actually current and copy them to our
8:27 tiered storage drive here the idea is that i want to bring over more than
8:32 would comfortably fit in the cache and then i want the editors during our test
8:37 to open up first a whole set of completely different projects from each
8:40 other then close those open up a new set
8:44 go back open the other ones and evaluate the performance for themselves
8:49 so we're back and we settled into a steady transfer speed of anywhere from
8:54 200 to 300 plus megabytes a second
8:58 and all of our projects are over on here so
9:02 i guess all that's left for me to do then is go interrupt the editing team
9:08 no no no so i just need to know if you guys all
9:12 hit this and try to edit 8k footage at the same time
9:17 will you notice slow downs because i know some people are
9:21 sensitive to slow downs so here's the idea
9:26 i don't even know what a network share is this is the network share but i don't
9:30 know what that is uh so everyone kind of pick a project and
9:34 open it and start doing video editor things whatever those are work the way
9:37 you would normally work because full is going to be affected by like your system
9:41 as well so the idea
9:47 is that once it arrives it gets cached in the fast tier so i
9:53 just wanted to get a feel for what this is like ed are you opening a project oh
9:58 you're coding wait you're encoding it oh uh exporting probably isn't a very
10:03 stressful test okay what are you doing taren i uh
10:07 i don't know can you try doing some more editory things like pretend to edit this
10:11 i'm moving stuff around i you know keep pretending pretend to do your job
10:17 yes can you go back and play that sequence again
10:22 and see if it's any better how fast would you expect it to be
10:28 like is this normal or it's a little bit faster
10:33 it's not that way okay just hold on a second i haven't i
10:37 haven't given up on saving money yet so the thing i want to find out now is
10:42 if you guys keep doing stuff will it get any better
10:47 so i'm starting to come around to the realization that
10:51 maybe tiered storage solutions
10:54 aren't grab and go because in every environment you'd want
10:59 your tiers to behave in a completely different manner
11:04 like if there was some programmability in storage spaces where i could say okay
11:08 take all the newest files that come in and assume they are important
11:13 until such time as they do not get accessed for
11:17 a very long period or until such time as something new comes in and
11:22 bumps it out then it would be less likely for those
11:26 guys to open up a project and have it chug until
11:30 storage spaces goes oh this is important and can promote it from the slow
11:34 mechanical tier so it's possible that if i had a much much larger fast NVMe tier
11:40 we wouldn't run into this issue because this is really a very heavy load for
11:44 four hard drives and two ssds but the problem is that that would
11:48 require me to basically build that entire server without even knowing that
11:53 it would work so you are dropping i'm dropping 90 of
11:57 my frames can you rate your editing experience out of 10 for me here there's
12:01 like little moments of seven out of ten
12:04 but then overall it's more like a three out of ten sorry you're saying this is
12:09 better than waning a little bit maybe all right everyone just go back to wanik
12:13 unmap the drive throw it all in the garbage
12:22 so that's pretty much it i am
12:26 coming to the the the acceptance stage
12:30 in the in the stages of grief here um
12:35 where i'm realizing that if we want
12:38 more capacity of our high performance network storage
12:42 i'm just going to have to pull out the wallet and make it happen
12:48 thankfully to take some of the sting out of it is our sponsor for this video
12:53 squarespace with squarespace you can make your own beautiful website without
12:58 the hassle they're all in one platform makes it easy to get up and running
13:02 quickly and they've got all kinds of great features from their logo designer
13:06 to their built-in e-commerce to the ability to publish in apple news format
13:10 so that your blog will be available immediately to millions of potential
13:14 readers worldwide and they offer all kinds of support so you can build the
13:18 best website you can build whether it's their seminars or contacting their 24 7
13:24 tech support via live chat and email so don't take my word for it go try out
13:28 squarespace at the link in the video description then when you decide to sign
13:32 up for real go to squarespace.com LTT and you can get 10 off your first
13:37 purchase so thanks for watching guys if you
13:40 disliked this video you can hit that button but if you liked it hit like get
13:43 subscribed or maybe consider checking out where to buy the stuff we featured at the link in the video description
13:48 also down there is our merch store which has cool shirts
13:53 it's winter like this one and our community forum which you should totally
13:57 join