The Inverted Pyramid of Doom Challenge
-
@hutchingsp sorry that it was so long for me to address the post. When you first posted it seemed reasonable and we did not have any environment of our own that exactly addressed the scale and needs that you have. But for the past seven months, we've been running on a Scale cluster, first a 2100 and now a 2100/2150 hybrid and that addresses every reason that I feel that you were avoiding RLS and addresses them really well.
The concern that I would have with your design is not reliability, the HUS SANs are very reliable. But they are far from free and far from easy to use. That cost is not cheap. You are in great shape that you can do direct connections (DAS) rather than SAN for better reliability and nominally better performance with much lower cost, but at three hosts plus the SAN, I can only imagine that you had to spend a lot more for that complex and limited setup than you would have with RLS that would not have the issues that you feel that it would.
Of course to really know we'd have to compare real world numbers. And compare soft benefits. How well suited is the HUS plus three node design to scaling up? If you add one more node, do you need switches? Is it two more nodes? With the RLS design, in our situation at least, we can add many additional nodes and scale up transparently gaining capacity and IOPS performance and throughput.
I don't have enough details to compare the final decision that you made compared to some other solutions that are out there. But in the real world, the reasons that you listed as to why you felt RLS was not an option for you are demonstrably not concerns in our environment. It's not a theoretical case, but one that we run every day.
-
@hutchingsp said in The Inverted Pyramid of Doom Challenge:
The HUS is 99.999% uptime rated - the environment it's in is not, so the HUS is not the thing I should be worrying about in terms of reliability.
We get that from normal servers. If the HUS isn't saving costs, nor improving reliability, what is the benefit? I'm shocked that the HUS is only five nines. Not that that is the deciding factor, but with all of your concerns about firmware, failing of 10TB workloads and such... why did that not rule out the HUS? It seems like either the HUS isn't "up" enough to justify the cost and/or your other downtimes are so great that it doesn't matter anywhere. I'm just curious as to the cost benefit of the HUS over local storage given that your HUS isn't getting a significant improvement in uptime and the downtime that it does get is hidden in the "background noise" or your environment's outages?
-
In the scenario above, using the SAN here is much like a lab scenario. You don't care about the reliability, it's "reliable enough" and it is recognized that reliability is not the concern. So that is much like a lab environment. What I am unclear about, though is without that reliability goal, what goal did the SAN address? It seems like it is both slower and much more expensive than local storage would be. So cost and performance don't seem like they could be the drivers. And while reliability was listed as not a key factor, it was mentioned in the list as a key factor, which seems to go against the final conclusion.
-
Actually had one come up today where a company had thirteen hosts and was using a SAN. Of course, the problem with this isn't that the design is wrong but that it totally fits the guidelines of when you should use a SAN. So in no way does it go against the stated guidance on SAN and IPOD choices. It's an enterprise SAN with more than thirteen attached nodes (and more being added) so it is well within the range of scale where we expect an IPOD/SAN combo to be totally logical and valid.
https://community.spiceworks.com/topic/1741842-migrating-boot-from-san-hosts
The real question is "is this actually an SMB that just has huge needs, or is it a larger business?" But in either case, it's a good example of where a SAN actually makes total sense (but isn't a requirement, other viable options at that size, too), but does so meeting the established, traditional guidelines (more than a dozen nodes on the same shared storage pool) so is not an exception or revelation in any way.
-
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
We have this now and we use the same capacity with replicated local disks as you would with a SAN with RAID 10. Are you using RAID 6 or something else to get more capacity from the SAN than you can with RLS? We aren't wasting any capacity having the extra redundancy and reliability of the local disks with RAIN.
With HDT pools you could have Tier 0 be RAID 5 SSD, RAID 10 for 10K's in the middle Tier, and RAID 6 for NL-SAS with sub-lun block tiering across all of that. Replicated local you generally can't do this (or you can't dynamically expand individual tiers). Now as 10K drives make less sense (hell magnetic drives make less sense) the cost benefits of a fancy tiering system might make less sense. Then again I see HDS doing tiering now between their custom FMD's regular SSD's, and NL's in G series deployments so there's still value in having a big ass array that can do HSM.
-
@scottalanmiller His previous RAIN system (HP Store Virtual) requires local RAID. So your stuck with nested RAID 5's (Awful IO performance) or nested RAID 10 (awful capacity capabilities, but great residency). Considering the HDS has 5 nines already though kinda moot.
-
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
@hutchingsp said in The Inverted Pyramid of Doom Challenge:
In the end I ruled it out because it introduces more complexity than I wanted...
I'm confused here. How does RLS introduce more complexity? Our RLS system is so simple I couldn't possible install any enterprise SAN on the market as easily. It takes literally no knowledge or setup at all. We are using Scale's RLS hyperconverged system and I literally cannot fathom it being easier today. Just making a LUN would be more complex than using RLS alone. Just needing to know that you need to make a LUN is more complex. With the RLS that we have, you just "use it". If you want more power, you can choose to manage your performance tiering manually, but there is no need to if you don't want to as it does it automatically for you out of the box.
To be fair in comparison he could have deployed Hitachi Unified Compute (Their CI stack) and gotten the same experience (basically someone builds out the solution for you and gives you pretty automated tools to abstract the complexity away, as well as have API drive endpoints for provisioning against etc). This isn't an argument for an architecture (and I like HCI, I REALLY do) this is an argument about build vs. buy your making. HCI with VSA's can be VERY damn complicated. HCI can be very simple (especially when it's built by someone else). CI can do this too.
-
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
@hutchingsp sorry that it was so long for me to address the post. When you first posted it seemed reasonable and we did not have any environment of our own that exactly addressed the scale and needs that you have. But for the past seven months, we've been running on a Scale cluster, first a 2100 and now a 2100/2150 hybrid and that addresses every reason that I feel that you were avoiding RLS and addresses them really well.
The other issue with Scale (and this is no offense to them) or anyone in the roll your own hypervisor game right now is there are a number of verticle applications that the business can not avoid, that REQUIRE specific hypervisors, or storage be certified. To get this certification you have to spend months working with their support staff to validate capabilities (performance, availability, predictability being one that can strike a lot of hybrid or HCI players out) as well as commit to cross engineering support with them. EPIC EMR (and the underlying Cachè database), application from industrial control systems from Honeywell and all kinds of others.
This is something that takes time, it takes customers asking for it, it takes money, and it takes market clout. I remember when even basic Sage based applications refused to support virtualization at all (then Hyper-V). It takes time for market adoption and even in HCI there are still some barriers (SAP is still dragging their feat on HANA certifications for HCI). At the end of the day customer choice is great, and if you can be a trailblazer and risk support to help push vendors to be more opened minded (That's great) but not everyone can do this.
There are other advantages to his design over a HCI design. If he has incredibly data heavy growth in his environment he doesn't have to add hosts. As licensing for Microsoft applications stacks (datacenter, SQL etc) are tied to CPU Core's here in the near future adding hosts to add storage can be come rather expensive if you don't account for it properly. Now you could mount external storage to the cluster to put the growing VM's on, but I'm not sure if Scale Supports that? He also within the HUS can either grow existing pools, add new pools (maybe a dedicated cold Tier 3) or PIN LUN's to a given tier (Maybe put a database always in flash). There's a lot of control here of storage costs and performance (if you have patience to manage it. Sadly no VVOLs support coming to the old HUS's.
-
At the time of going the HUS route the HCI solutions were limited.
We looked at a couple of the big names, pricing was insane though I'd happily have a rack of either if money wasn't a factor in any of the lifecycle beyond just the initial purchase price.
That left a whole bunch of "roll your own" solutions, Scale at the time didn't seem to scale plus we're comfortable on ESXi, could we change? Yes of course, but then you're migrating to a hypervisor with a more limited support ecosystem however rock solid that hypervisor may be so I'm not sure what we actually gain there?
Starwind and other solutions are a bit too roll-your-own as was, I felt, StoreVirtual, in that if it all went wrong you're possibly stuck between the hardware vendor, the switch vendor, the OS vendor, the HCI stack vendor and so on which is great when you've saved a few $$$ but not so good when your business is down and nobody wants to own the issue.
Scale, Simplivity, Nutanix etc. are (or were when we purchased, maybe it's changed) all designed on the principle you scale compute and storage linearly and we don't which means that if/when I needed to add another TB of capacity I'd either be bolting on NAS and stuff to a HCI solution, or I'd be buying additional nodes giving me compute and storage and licensing costs when really all I needed was a disk tray and a couple of drives which I can do with the HDS.
In short it's simple and it's reliable and from the numbers at the time I know we didn't pay massively more than anything I could have rolled ourselves and ended up with a lot more complexity for very little gain.
That said, when the replacement time comes I suspect the HCI and VSA options will have matured a lot from where they were 2 years ago so there will be some more interesting options on the table - but I see that as being more in terms of things like stretching clusters than chasing 9's from a pile of kit all sitting in the same rack.
-
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
to hit checkboxes while resting on fragile, single point of failure storage (SAN, NAS, DAS, etc.)
As I've heard many times, even once from one of my own employees who argued with me about this "but how often do SANs/NASes/etc fail? Basically never, so it's not a single point of failure, and it makes it easier to move from one VM host to another."
I'm still baffled by this, and actually have seen them fail and take all customer data with them. What in God's name makes people seem to believe that you're more likely to have a motherboard or power supply fail in a server (since they don't use the disks) than actual harddisks. For well over a year we've had a no hire policy during interview if someone describes inverted pyramid as a proper way to set something up,. Yes, they can learn to do otherwise, but:
- I don't want to teach them nor do I want my employees to waste their time teaching them something they should have figured out themselves.
- If they just thought about it logically for even a second, they'd see how it's probably a stupid idea to put all your eggs in one basket.
I have no regrets. At no point in time was "needing the space" a good enough reason, unless the SAN was for one VM host or something like that possibly, but these days that kind of thing is rare anyway, at least in the environments for our customers, and our own.
-
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
We have this now and we use the same capacity with replicated local disks as you would with a SAN with RAID 10. Are you using RAID 6 or something else to get more capacity from the SAN than you can with RLS? We aren't wasting any capacity having the extra redundancy and reliability of the local disks with RAIN.
With HDT pools you could have Tier 0 be RAID 5 SSD, RAID 10 for 10K's in the middle Tier, and RAID 6 for NL-SAS with sub-lun block tiering across all of that. Replicated local you generally can't do this (or you can't dynamically expand individual tiers). Now as 10K drives make less sense (hell magnetic drives make less sense) the cost benefits of a fancy tiering system might make less sense. Then again I see HDS doing tiering now between their custom FMD's regular SSD's, and NL's in G series deployments so there's still value in having a big ass array that can do HSM.
We have only two tiers, but they can be dynamically expanded. Any given node can be any mix of all slow tier, all fast tier or a blend. There is a standard just because it's pre-balanced for typical use, but nothing ties it to that.
-
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
@scottalanmiller His previous RAIN system (HP Store Virtual) requires local RAID. So your stuck with nested RAID 5's (Awful IO performance) or nested RAID 10 (awful capacity capabilities, but great residency). Considering the HDS has 5 nines already though kinda moot.
Oh I know that some systems can be limited and lack functionality. But that doesn't imply that it is a limitation of the concept or a limitation of available implementations. It may be (or have been) but is not currently. HP VSA isn't bad, but I rarely think of that as the "go to" implementation of a RAIN system.
-
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
There are other advantages to his design over a HCI design. If he has incredibly data heavy growth in his environment he doesn't have to add hosts. As licensing for Microsoft applications stacks (datacenter, SQL etc) are tied to CPU Core's here in the near future adding hosts to add storage can be come rather expensive if you don't account for it properly. Now you could mount external storage to the cluster to put the growing VM's on, but I'm not sure if Scale Supports that?
I don't know about every vendor obviously, but I know that with our Scale we can swap out existing drives for larger ones or faster ones (as long as we don't lose more capacity than we need, obviously) if we need to rebalance our capacity. So some of our nodes are 1.8TB disks, for example. If our needs changed and we needed to go to 6TB disks, we could just swap them out and grow that way (in place.)
If we needed more growth than "bigger" drives would account for and/or we needed to increase performance or fault tolerance at the same time (bigger drives means more capacity bounded by the same IOPS which can be a problem) we can add "storage only" nodes. They can't run VMs but will grow the storage pool. These are rare to use and very few people talk about them because you can also just buy a normal compute node with storage and tell your workloads not to go there to solve the licensing issues 99% of the time. That way you get more performance and "fewer eggs in the same baskets" out of the investment. But you can balance things as needed.
I don't know how many players do that. I'm pretty sure that Starwinds does. I know that rolling your own with Gluster or CEPH (far from ideal) will do that. It's at least a semi-standard feature in the RLS space these days.
DRBD does not have this yet. I've not tried HAST yet, on my long list of things to put into the lab, but I'm pretty sure that it does not either. But you should really think of those only as two node solution tools, not scale out ones.
-
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
The other issue with Scale (and this is no offense to them) or anyone in the roll your own hypervisor game right now is there are a number of verticle applications that the business can not avoid, that REQUIRE specific hypervisors, or storage be certified. To get this certification you have to spend months working with their support staff to validate capabilities (performance, availability, predictability being one that can strike a lot of hybrid or HCI players out) as well as commit to cross engineering support with them.
This is always tough and is certainly a challenge to any product. It would be interesting to see a survey of just how often this becomes an issue and how it is addressed in different scenarios. From my perspective, and few companies can do this, it's a good way to vet potential products. Any software vendor that needs to know what is "under the hood" isn't ready for production at all. They might need to specify IOPS or resiliency or whatever, sure. But caring about the RAID level used, that it is RAID or RAIN, what hypervisor is underneath the OS that they are given - those are immediate show stoppers, no vendor with those kinds of artificial excuses to not provide support are show the door. Management should never even know that the company exists as they are not viable options and not prepared to support their products. Whether it is because they are incompetent, looking for kickbacks or just making any excuse to not provide support does not matter, it's something that a business should not be relying on for production.
-
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
We have this now and we use the same capacity with replicated local disks as you would with a SAN with RAID 10. Are you using RAID 6 or something else to get more capacity from the SAN than you can with RLS? We aren't wasting any capacity having the extra redundancy and reliability of the local disks with RAIN.
With HDT pools you could have Tier 0 be RAID 5 SSD, RAID 10 for 10K's in the middle Tier, and RAID 6 for NL-SAS with sub-lun block tiering across all of that. Replicated local you generally can't do this (or you can't dynamically expand individual tiers). Now as 10K drives make less sense (hell magnetic drives make less sense) the cost benefits of a fancy tiering system might make less sense. Then again I see HDS doing tiering now between their custom FMD's regular SSD's, and NL's in G series deployments so there's still value in having a big ass array that can do HSM.
We have only two tiers, but they can be dynamically expanded. Any given node can be any mix of all slow tier, all fast tier or a blend. There is a standard just because it's pre-balanced for typical use, but nothing ties it to that.
The other advantage of having tiers with different raid levels etc, is he can use RAID 6 NL-SAS for ICE cold data, and RAID 5/10 for higher tiers for better performance. Only a few HCI solutions today do true always on erasure codes in a way that isn't murderous to performance during rebuilds. (GridStore, VSAN, ?)
-
Cost. Mirroring has a 2/3x overhead for FTT=1/2 while erasure codes can get that much much lower (IE half as many drives for FTT=2, or potentially less depending on stripe width). As we move to all flash in HCI (it coming) the IO/latency overhead for Erasure Codes and Dedupe/Compression becomes negligible. This is a competitive gap between several different solutions right now in that space.
-
When your adding nodes purely for capacity this carries other non-visible costs. Power draw (A Shelf on that HUS draws a lot less than a server), scale out systems consume more ports (and while this benefits throughput, and network ports are a LOT cheaper this is more structured cabling, more ports to monitor for switch monitoring licensing etc).
At small scale none of this matters that much (OPEX benefits in labor and support footprint trump these other costs). At mid/large scale this stuff adds up...
-
-
@scottalanmiller 1.8TB drives if they are 10K are 2.5'' 6TB are 3.5''. If they can stuff a 3.5'' drive in a 2.5'' bay I'd be impressed.
The reality of 10K drives is the roadmap is dead. I don't expect to see anything over 1.8TB and in reality because those are 512E/4KN block drives anyone with any legacy OS's ends up stuck with 1.2TB drives more often than not if they don't want weird performance issues.
(Fun fact, Enterprise Flash drives are ALL 512E for 4KN back ends, but it doesn't matter because they have their own write buffers that absorb and re-order the writes to prevent any amplification).
Storage nodes are not commonly used, but largely because the vendors effectively charge you the same amount for them (At least the pricing on Nutanix storage only nodes wasn't that much of a discount). Outside of licensing situations no one would ever buy them up front (they would have right sized the cluster design from the start). In reality they something you kinda get forced into buying (You can't add anyone else's storage to the cluster).
I get the opex benefits of CI and HCI appliances, but the fact that you've completely frozen any flexibility on servers and storage comes at a cost, and that's lack of control by the institution on expansion costs.
-
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
@scottalanmiller 1.8TB drives if they are 10K are 2.5'' 6TB are 3.5''. If they can stuff a 3.5'' drive in a 2.5'' bay I'd be impressed.
It's all 3.5" bays, universally. Scale has no 2.5" bay offerings.
-
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
I get the opex benefits of CI and HCI appliances, but the fact that you've completely frozen any flexibility on servers and storage comes at a cost, and that's lack of control by the institution on expansion costs.
I wouldn't call it "completely frozen." You can, at least on the Scale and some other vendors systems, mix and match storage-only and compute nodes, you can get different nodes with different capacity or performance drive options, you can get different processor sizes (core count) and speeds and even different processor counts, different memory configurations, etc. Sure, if you want to mix and match AMD64 with Power8 or Sparc64 systems you are out of luck. That's a true limitations. But are you running into many customers that want to do that? It exists, but it is very, very rare. Especially in companies small enough to fit into the scale limitations of most HCI solutions. Outside of leaving the AMD64 world, which few SMBs do, there are very few limitations.
-
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
Storage nodes are not commonly used, but largely because the vendors effectively charge you the same amount for them (At least the pricing on Nutanix storage only nodes wasn't that much of a discount).
Because those nodes need CPU and RAM as well, there is only so much room for discounts. However, when doing this often you are doing it because, for example, you want the storage and failover of three nodes but only want Windows DC licensing on two nodes. So you pay for two nodes with dual high end Intel Xeons with lots of cores and high clock speeds, gobs of RAM, etc. Then the storage node is just a single, low end proc and a very small amount of RAM. So while the storage only "discount" might be very small, the amount that you have to pay less for the node compared to the compute nodes can be massive.
-
@scottalanmiller said in The Inverted Pyramid of Doom Challenge:
@John-Nicholson said in The Inverted Pyramid of Doom Challenge:
The other issue with Scale (and this is no offense to them) or anyone in the roll your own hypervisor game right now is there are a number of verticle applications that the business can not avoid, that REQUIRE specific hypervisors, or storage be certified. To get this certification you have to spend months working with their support staff to validate capabilities (performance, availability, predictability being one that can strike a lot of hybrid or HCI players out) as well as commit to cross engineering support with them.
This is always tough and is certainly a challenge to any product. It would be interesting to see a survey of just how often this becomes an issue and how it is addressed in different scenarios. From my perspective, and few companies can do this, it's a good way to vet potential products. Any software vendor that needs to know what is "under the hood" isn't ready for production at all. They might need to specify IOPS or resiliency or whatever, sure. But caring about the RAID level used, that it is RAID or RAIN, what hypervisor is underneath the OS that they are given - those are immediate show stoppers, no vendor with those kinds of artificial excuses to not provide support are show the door. Management should never even know that the company exists as they are not viable options and not prepared to support their products. Whether it is because they are incompetent, looking for kickbacks or just making any excuse to not provide support does not matter, it's something that a business should not be relying on for production.
This right here makes no sense to me. You are ok with recommending infrastructure that can ONLY be procured from a single vendor for all expansions and has zero cost control of support renewal spikes, hardware purchasing, software purchasing (Proprietary hypervisor only sold with hardware), but you can't buy a piece of software that can run on THOUSANDS of different hardware configurations and more than 1 bare metal platform?
In Medicine for EMR's Caché effectively controls the database market for anyone with multiple branches with 300+ beds and one of every service. (Yes there is All Scripts that runs on MS SQL, and no it doesn't scale and is only used for clinics and the smallest hospitals as Philip will tell you). If you tell the chief of medicine you will only offer him tools that will not scale to his needs you will (and should) get fired. There are people who try to break out from the stronghold they have (MD Anderson who has a system that is nothing more than a collection of PDF's) but its awful, and doctors actively choose not to work in these hospitals because the IT systems are so painful to use (You can't actually look up what medicines someone is on, you have to click through random PDF's attached to them or find a nurse). The retraining costs that would be required to fulfill an IT mandate of "can run on any OS or hypervisors" to migrate this platform (or many major EMR's) are staggered. IT doesn't have this much power even in the enterprise. Sometimes the business driver for a platform outweighs the loss of stack control, or conformity of infrastructure (I"m sure the HFT IT guys had this drilled into their head a long time ago). This is partly the reason many people still quietly have a HP-UX, or AS400 in the corner still churning their ERP..
I agree that you should strongly avoid lock in where possible, but when 99% of the other community is running it on X and your the 1% that makes support calls a lot more difficult, not just because they are not familiar with your toolset. A lot of these products have cross engineering escalation directly into the platforms they have certified. We have lock-in on databases for most application stacks (and live with it no matter how many damn yachts we buy Larry). The key things are:
-
Know the costs going in. Don't act surprised when you buy software for 1 million and discover you need 500K worth of complementary products and hardware to deploy it.
-
Know what parts you can swap if they fail to deliver. (Hardware, support, OS, Database, Hypervisor) and be comfortable with any with reduced choice, or no choice. Different people may need different levels of support for each.
-
Also know what your options for the platform for hosted or OPEX non-hardware offerings are (IE can I replicate to a multi-tenant DR platform to make DR a lowered Opex).
-
