You are currently browsing the tag archive for the ‘tiering’ tag.
I have a customer who just recently upgraded their EMC Celerra NS480 Unified Storage Array (based on Clariion CX4-480) to FLARE30 and enabled FASTCache across the array, as well as FASTVP automated tiering for a large amount of their block data. Now that it’s been configured and the customer has performed a large amount of non-disruptive migrations of data from older RAID groups and VP pools into the newer FASTVP pool, including thick-to-thin conversions, I was able to get some performance data from their array and thought I’d share these results.
This is Real-World data
This is NOT some edge case where the customer’s workload is perfect for FASTCache and FASTVP and it’s also NOT a crazy configuration that would cost an arm and a leg. This is a real production system running in a customer datacenter, with a few EFDs split between FASTCache and FASTVP and some SATA to augment capacity in the pool for their existing FC based LUNS. These are REAL results that show how FASTVP has distributed the IO workload across all available disks and how a relatively small amount of FASTCache is absorbing a decent percentage of the total array workload.
This NS480 array has nearly 480 drives in total and has approximately 28TB of block data (I only counted consumed data on the thin LUNs) and about 100TB of NAS data. Out of the 28TB of block LUNs, 20TB is in Virtual Pools, 14TB of which is in a single FASTVP Pool. This array supports the customers’ ERP application, entire VMWare environment, SQL databases, and NAS shares simultaneously.
In this case FASTCache has been configured with just 183GB of usable capacity (4 x 100GB EFD disks) for the entire storage array (128TB of data) and is enabled for all LUNs and Pools. The graphs here are from a 4 hour window of time after the very FIRST FASTVP re-allocation completed using only about 1 days’ worth of statistics. Subsequent re-allocations in the FASTVP pool will tune the array even more.
First, let’s take a look at the array as a whole, here you can see that the array is processing approximately ~10,000 IOPS through the entire interval.
FASTCache is handling about 25% of the entire workload with just 4 disks. I didn’t graph it here but the total array IO Response time through this window is averaging 2.5 ms. The pools and RAID Groups on this array are almost all RAID5 and the read/write ratio averages 60/40 which is a bit write heavy for RAID5 environments, generally speaking.
If you’ve done any reading about EMC FASTCache, you probably know that it is a read/write cache. Let’s take a look at the write load of the array and see how much of that write load FASTCache is handling. In the following graph you can see that out of the ~10,000 total IOPS, the array is averaging about 2500-3500 write IOPS with FASTCache handling about 1500 of that total.
That means FASTCache is reducing the back-end writes to disk by about 50% on this system. On the NS480/CX4-480, FASTCache can be configured with up to 800GB usable capacity, so this array could see higher overall performance if needed by augmenting FASTCache further. Installing and upgrading FASTCache is non-disruptive so you can start with a small amount and upgrade later if needed.
FASTVP and FASTCache Together
Next, we’ll drill down to the FASTVP pool which contains 190 total disks (5 x EFD, 170 x FC, and 15 x SATA). There is no maximum number of drives in a Virtual Pool on FLARE30 so this pool could easily be much larger if desired. I’ve graphed the IOPS-per-tier as well as the FASTCache IOPS associated with just this pool in a stacked graph to give an idea of total throughput for the pool as well as the individual tiers.
The pool is servicing between 5,000 and 8,000 IOPS on average which is about half of the total array workload. In case you didn’t already know, FASTVP and FASTCache work together to make sure that data is not duplicated in EFDs. If data has been promoted to the EFD tier in a pool, it will not be promoted to FASTCache, and vise-versa. As a result of this intelligence, FASTCache acceleration is additive to an EFD-enabled FASTVP pool. Here you can see that the EFD tier and FASTCache combined are servicing about 25-40% of the total workload, the FC tier another 40-50%, and the SATA tier services the remaining IOPS. Keep in mind that FASTCache is accelerating IO for other Pools and RAID Group LUNs in addition to this one, so it’s not dedicated to just this pool (although that is configurable.)
FASTVP IO Distribution
Lastly, to illustrate FASTVP’s effect on IO distribution at the physical disk layer, I’ve broken down IOPS-per-spindle-per-tier for this pool as well. You can see that the FC disks are servicing relatively low IO and have plenty of head room available while the EFD disks, also not being stretched to their limits, are servicing vastly more IOPS per spindle, as expected. The other thing you may have noticed here is that the EFDs are seeing the majority of the workload’s volatility, while the FC and SATA disks have a pretty flat workload over time. This illustrates that FASTVP has placed the more bursty workloads on EFD where they can be serviced more effectively.
Hopefully you can see here how a very small amount of EFDs used with both FASTCache and FASTVP can relieve a significant portion of the workload from the rest of the disks. FASTCache on this system adds up to only 0.14% of the total data set size and the EFD tier in the FASTVP pool only accounts for 2.6% of the total dataset in that pool.
What do you think of these results? Have you added FASTCache and/or FASTVP to your array? If so, what were your results?
Some customers are afraid of thin provisioning…
Practically every week I have discussions with customers about leveraging thin provisioning to reduce their storage costs and just as often the customer pushes back worried that some day, some number of applications, for some reason, will suddenly consume all of their allocated space in a short period of time and cause the storage pool to run out of space. If this was to happen, every application using that storage pool will essentially experience an outage and resolving the problem requires allocating more space to the pool, migrating data, and/or deleting data, each of which would take precious time and/or money. In my opinion, this fear is the primary gating factor to customers using thin provisioning. Exacerbating the issue, most large organizations have a complex procurement process that forces them to buy storage many months in advance of needing it, further reducing the usefulness of thin provisioning. The IT organization for one of my customers can only purchase new storage AFTER a business unit requests it and approved by senior management; and they batch those requests before approving a storage purchase. This means that the business unit may have to wait months to get the storage they requested.
This same customer recently purchased a Symmetrix VMAX with FASTVP and will be leveraging sub-LUN tiering with SSD, FC, and SATA disks totaling over 600TB of usable capacity in this single system. As we began design work for the storage array the topic of thin provisioning came up and the same fear of running out of space in the pool was voiced. To prevent this, the customer fully allocates all LUNs in the pool up front which prevents oversubscription. It’s an effective way to guarantee performance and availability but it means that any free space not used by application owners is locked up by the application server and not available to other applications. If you take their entire environment into account with approximately 3PB of usable storage and NO thin provisioning, there is probably close to $1 million in storage not being used and not available for applications. If you weigh the risk of an outage causing the loss of several million dollars per hour of revenue, the customer has decided the risks outweigh the potential savings. I’ve seen this decision made time and again in various IT shops.
Sub-LUN Tiering pushes the costs for growth down
I previously blogged about using cloud storage for block storage in the form of Cirtas BlueJet and how it would not be to much of a stretch to add this functionality to sub-LUN tiering software like EMC’s FASTVP to leverage cloud storage as a block storage tier as shown in this diagram.
Let’s first assume the customer is already using FASTVP for automated sub-LUN tiering on a VMAX. FASTVP is already identifying the hot and cold data and moving it to the appropriate tier, and as a result the lowest tier is likely seeing the least amount of IOPS per GB. In a VMAX, each tier consists of one or more virtual provisioned pools, and as the amount of data stored on the array grows FASTVP will continually adjust, pushing the hot data up to higher tiers and cold data down to the lower tiers The cold data is more likely to be old data as well so in many cases the data sort of ages down the tiers over time and its the old/least used portion of the data that grows. Conceptually, the only tier you may have to expand is the lowest (ie: SATA) when you need more space. This reduces the long term cost of data growth which is great. But you still need to monitor the pools and expand them before they run out of space, or an outage may occur. Most storage arrays have alerts and other methods to let you know that you will soon run out of space.
Risk-Free Thin Provisioning
What if the storage array had the ability automatically expand itself into a cloud storage provider, such as AT&T Synaptic, to prevent itself from running out of space? Technically this is not much different from using the cloud as a tier all it’s own but I’m thinking about temporary use of a cloud provider versus long term. The cloud provider becomes a buffer for times when the procurement process takes too long, or unexpected growth of data in the pool occurs. With an automated tiering solution, this becomes relatively easy to do with fairly low impact on production performance. In fact, I’d argue that you MUST have automated tiering to do this or the array wouldn’t have any method for determining what data it should move to the cloud. Without that level of intelligence, you’d likely be moving hot data to the cloud which could heavily impact performance of the applications.
Once the customer is able to physically add storage to the pool to deal with the added data, the array would auto-adjust by bringing the data back from the cloud freeing up that space. The cloud provider would only charge for the transfer of data in/out and the temporary use of space. Storage reduction technologies like compression and de-duplication could be added to the cloud interface to improve performance for data stored in the cloud and reduce costs. Zero detect and reclaim technologies could also be leveraged to keep LUNs thin over time as well as prevent the movement of zero’d blocks to the cloud.
Using cloud storage as a buffer for thin provisioning in this way could reduce the risk of using thin provisioning, increasing the utilization rate of the storage, and reducing the overall cost to store data.
What do you think? Would you feel better about oversubscribing storage pools if you had a fully automated buffer, even if that buffer cost some amount of money in the event it was used?
I came across this press release today from a company that I wasn’t familiar with and immediately wanted more information. Cirtas Systems has announced support for Atmos-based clouds, including AT&T Synaptic Storage. Whenever I see these types of announcements, I read on in hopes of seeing real fiber channel block storage leveraging cloud-based architectures in some way. So far I’ve been a bit disappointed since the closest I’ve seen has been NAS based systems, at best including iSCSI.
Cirtas BlueJet Cloud Storage Controller is pretty interesting in its own right though. It’s essentially an iSCSI storage array with a cache and a small amount of SSD and SAS drives for local storage. Any data beyond the internal 5TB of usable capacity is stored in “the cloud” which can be an onsite Private Cloud (Atmos or Atmos/VE) and/or a Public Cloud hosted by Amazon S3, Iron Mountain, AT&T Synaptic, or any Atmos-based cloud service provider.
The neat thing with BlueJet is that it leverages a ton of the functionality that many storage vendors have been developing recently such as data de-duplication, compression, some kind of block level tiering, and space efficient snapshots to improve performance and reduce the costs of cloud storage. It seems that pretty much all of the local storage (SAS, SSD, and RAM) is used as a tiered cache for hot data. This gives users and applications the sense of local SAN performance even while hosting the majority of data offsite.
While I haven’t seen or used a BlueJet device and can’t make any observations about performance or functionality, I believe this sort of block->cloud approach has pretty significant customer value. It reduces physical datacenter costs for power and cooling, and it presents some rather interesting disaster recovery opportunities.
Similar to how Compellent’s signature feature, tiered block storage, has been added to more traditional storage arrays, I think modified implementations of Cirtas’ technology will inevitably come from the larger players, such as EMC, as a feature in standard storage arrays. If you consider that EMC Unified Storage and EMC Symmetrix VMAX both have large caches and block- level tiering today, it’s not too much of a stretch to integrate Atmos directly into those storage systems as another tier. EMC already does this for NAS with the EMC File Management Appliance.
I can imagine leveraging FASTCache and FASTVP to tier locally for the data that must be onsite for performance and/or compliance reasons and pushing cold/stale blocks off to the cloud. Additionally, adding cloud as a tier to traditional storage arrays allows customers to leverage their existing investment in Storage, FC/FCoE networks, reporting and performance trending tools, extensive replication options available, and the existing support for VMWare APIs like SRM and VAAI.
With this model, replication of data for disaster recovery/avoidance only needs to be done for the onsite data since the cloud data could be accessed from anywhere. At a DR site, a second storage system connects to the same cloud and can access the cold/stale data in the event of a disaster.
Another option would be adding this functionality to virtualization platforms like EMC VPLEX for active/active multi-site access to SAN data, while only needing to store the majority of the company’s data once in the cloud for lower cost. Customers would no longer have to buy double the required capacity to implement a disaster recovery strategy.
I’m eagerly awating the implementation of cloud into traditional block storage and I can see how some vendors will be able to do this easily, while others may not have the architecture to integrate as easily. It will be interesting to see how this plays out.
I’ve been having some fun discussions with one of my customers recently about how to tackle various application problems within the storage environment and it got me thinking about the value of having “options”. This customer has an EMC Celerra Unified Storage Array that has Fiber Channel, iSCSI, NFS, and CIFS protocols enabled. This single storage system supports VMWare, SQL, Web, Business Intelligence, and many custom applications.
The discussion was specifically centered on ensuring adequate storage performance for several different applications, each with a different type of workload…
1.) Web Servers – Primarily VMs with general-purpose IO loads and low write ratios.
2.) SQL Servers – Physical and Virtual machines with 30-40% write ratios and low latency requirements.
3.) Custom Application – A custom application database with 100% random read profiles running across 50 servers.
The EMC Unified solution:
EMC Storage already sports virtual provisioning in order to provision LUNs from large pools of disk to improve overall performance and reduce complexity. In addition, QoS features in the array can be used to provide guaranteed levels of performance for specific datasets by specifying minimum and maximum bandwidth, response time, and IO requirements on a per-LUN basis. This can help alleviate disk contention when many LUNs share the same disks, as in a virtual pool. Enterprise Flash Drives (EFD) are also available for EMC Storage arrays to provide extremely high performance to applications that require it and they can coexist with FC and SATA drives in the same array. Read and write cache can also be tuned at an array and LUN level to help with specific workloads. With the updates to the EMC Unified Platform that I discussed previously, Sub-LUN FAST (auto tiering), and FAST Cache (EFD used as array cache) will be available to existing customers after a simple, non-disruptive, microcode upgrade, providing two new ways to tackle these issues.
So which feature should my customer use to address their 3 different applications?
Sub-LUN FAST (Fully Automated Storage Tiering)
Put all of the data into large Virtual Provisioning pools on the array, add a few EFD (SSD) and SATA disks to the mix and enable FAST to automatically move the blocks to the appropriate tier of storage. Over time the workload would even out across the various tiers and performance would increase for all of the workloads with much fewer drives, saving on power, floor space, cooling, and potentially disk cost depending on the configuration. This happens non-disruptively in the background. Seems like a no-brainer right?
For this customer, FAST helps the web server VMs and the general-purpose SQL databases where the workload is predominately read and much of the same data is being accessed repeatedly (high locality of reference). As long as the blocks being accessed most often are generally the same, day-to-day, automated tiering (FAST) is a great solution. But what if the workload is much more random? FAST would want to push all of the data into EFD, which generally wouldn’t be possible due to capacity requirements. Okay, so tiering won’t solve all of their problems. What about FAST Cache?
Exponentially increase the size of the storage array’s read AND write cache with EFD (SSD) disks. This would improve performance across the entire array for all “cache friendly” applications.
For this customer, increasing the size of write cache definitely helps performance for SQL (50% increase in TPM, 50% better response time as an example) but what about their custom database that is 100% random read? Increasing the size of read cache will help get more data into cache and reduce the need to go to disk for reads, but the more random the data, the less useful cache is. Okay, so very large caches won’t solve all of their problems. EFDs must be the answer right?
Forget SATA and FC disks; just use EFD for everything and it will be super fast!! EFD has extremely high random read/write performance, low latency at high loads, and very high bandwidth. You will even save money on power and cooling.
The total amount of data this customer is dealing with in these three applications alone exceeds 20TB. To store that much in EFD would be cost prohibitive to say the least. So, while EFD can solve all of this customer’s technical problems, they couldn’t afford to acquire enough EFD for the capacity requirements.
But wait, it’s not OR, it’s AND
The beauty of the EMC Unified solution is that you can use all of these technologies, together, on the same array, simultaneously.
In this customer’s case, we put FC and SATA into a virtual pool with FAST enabled and provision the web and general-purpose SQL servers from it. FAST will eventually migrate the least used blocks to SATA, freeing the FC disks for the more demanding blocks.
Next, we extend the array cache using a couple EFDs and FAST Cache to help with random read, sequential pre-fetching, and bursty writes across the whole array.
Finally, for the custom 100% random read database, we dedicate a few EFDs to just that application, snapshot the DB and present copies to each server. We disable read and write cache for the EFD backed volumes which leaves more cache available to the rest of the applications on the array, further improving total system performance.
Now, if and when the customer starts to see disk contention in the virtual pool that might affect performance of the general-purpose SQL databases, QoS can be tuned to ensure low response times on just the SQL volumes ensuring consistent performance. If the disks become saturated to the point where QoS cannot maintain the response time or the other LUNs are suffering from load generated by SQL, any of the volumes can be migrated (non-disruptively) to a different virtual pool in the array to reduce disk contention.
If you look at offerings from the various storage vendors, many promote large virtual pools, some also promote large caches of some kind, others promote block level tiering, and a few promote EFD (aka SSDs) to solve performance problems. But, when you are consolidating multiple workloads into a single platform, you will discover that there are weaknesses in every one of those features and you are going to wish you had the option to use most or all of those features together.
You have that option on EMC Unified.