We raid vRAID

I took a bit of time off to read through Violin’s vRAID technology because I realized that vRAID (other than Violin’s vXM architecture) is the other most important technology that differentiates Violin Memory from the other upstarts. I blogged at a high-level about Violin a few entries ago, and we are continuing Violin impressive entrance with a storage technology that have been around for almost 25 years – RAID. Incidentally, I found this picture of the original RAID paper (see below):

Has RAID evolved with solid state storage? Evidently, no, because I have not read of any vendors (so far) touting any RAID revolution in their solid state offerings. There has been a lot of negative talks about RAID, but RAID has been the cornerstone and the foundation of storage ever since the beginning. But with the onslaughts of very large capacity HDDs, the demands of packing more bits-per-inch and the insatiable needs for reliability, RAID is slowly beginning to show its age. Cracks in the armour, I would say. And there are many newer, slightly more refined versions of RAID, from the Network RAID-style of HP P4000 or the Dell EqualLogic, to the RAID-X of IBM XIV, to innovations of declustered RAID in Panasas. (Interestingly, one of the early founders of the actual RAID concept paper, Garth Gibson, is the founder of Panasas).

And the new vRAID from Violin-System doesn’t sway much from the good ol’ RAID, but it has been adapted to address the issues of Solid State Devices.

Solid State devices (notably NAND Flash since everyone is using them) are very different from the usual spinning disks of HDDs. They behave differently and pairing solid state devices with the present implementations of RAID could be like mixing oil and water. I am not saying that the present RAID cannot work with solid state devices, but has RAID adapted to the idiosyncrasies of Flash?

It is like putting an old crank shaft into a new car. It might work for a while, but in the long run, it could damage the car. Similarly, conventional RAID might have detrimental performance and availability impact with solid state devices. And we have hardly seen storage vendors coming up to say that their RAID technology has been adapted to the solid state devices that they are selling. This silence could likely mean that they are just adapting to market requirements and not changing their RAID codes very much to take advantage of Flash or other solid state storage for that matter. Violin Memory has boldly come forward to meet that requirement and vRAID is their answer.

Violin argues that there are bottlenecks at the external RAID controller or software RAID level as well as use of legacy disk drive interfaces. And this is indeed true, because this very common RAID implementation squeezes performance at the expense of the other components such as CPU cycles.

Furthermore, there are plenty of idiosyncrasies in Flash with things such as erase-first, then write mechanism. The nature of NAND Flash, unlike DRAM, requires a block to be erased first before a write to the block is allowed. It does not “modify” per se, where the operations of read-modify-write is often applied in parity-based RAIDs of 5 and 6. Because of this nature, it is more like read-erase-write, and when the erase of the block is occurring, the read operation is stalled. That is why most SSDs will have impressive read latency (in microseconds), but very poor writes (in milliseconds). Furthermore, the parity-based RAID’s write penalty, can further aggravate the situation when the typical RAID technology is applied to NAND Flash solid state storage.

As the blocks in the NAND Flash build up, the accumulation of read-erase-write will not only reduce the lifespan of the blocks in the NAND Flash, it will also reduce the IOPS to a state we called Normalized Steady State. I wrote about this in my blog, “Not all SSDs are the same” some moons ago. In my blog, SNIA Solid State Storage Performance Testing Suite (SSS-PTS), there were 3 distinct phases of a typical NAND Flash SSD:

  • Fresh of out the Box (FOB)
  • Transition
  • Steady State
This performance degradation is part of what vendors call “Write Cliff”, where there is a sudden drop in IOPS performance as the NAND Flash SSD ages. Here’s a graph that shows the performance drop.
Violin’s vRAID, implemented within its switched vXM architecture itself, and using proprietary high performance flash controllers and the flash-optimized vRAID technology, is able deliver sustained IOPS throughout the lifespan of the flash SSD, as shown below:
To understand vRAID we have to understand the building blocks of the Violin storage array. NAND Flash chips of 4GB are packed into a Flash Package of 8 giving it 32GB. And 16 of these 32GB Flash Package are then consolidated into a 512GB VIMM (Violin Inline Memory Module). The VIMM is the starting block and can be considered as a “disk”, since we are used to the concept of “disk” in the storage networking world. 5 of these VIMMs will create a RAID group of 4+1 (four data and one parity), giving the redundancy, performance and capacity similar to RAID-5.
The block size used is 4K block and this 4K block is striped across the RAID group with 1K pages each on each of the VIMMs in the RAID group. Each of this 1K page is managed independently and can be placed anywhere in any flash block in the VIMMs, and spread out for lowest possible latency and bandwidth. This contributes to the “spike free latency” of Violin Memory. Additionally, there is ECC protection within each 1K page to correct flash bit error.
To protect against metadata corruption, there is an additional, built-in RAID Check bit to correct the VIMM errors. Lastly, one important feature that addresses the read-erase-write weakness of NAND Flash, the vRAID ensures that the slow erases never block a Read or a Write. This architectural feature enable spike-free latency in mixed Read/Write environments.
Here’s a quick overview of Violin’s vRAID architecture:
I still feel that we need a radical move away from the traditional RAID and vRAID is moving in the right direction to evolve RAID to meet the demands of the data storage market. Revolutionary and radical it may not be, but then again, is the market ready for anything else?
As I said, so far Violin is the only all-Flash vendor that has boldly come forward to meet the storage latency problem head-on, and they have been winning customers very quickly. Well done!

Not all SSDs are the same

Happy Lunar New Year! The Chinese around world has just ushered in the Year of the Water Dragon yesterday. To all my friends and family, and readers of my blog, I wish you a prosperous and auspicious Chinese New Year!

Over the holidays, I have been keeping up with the progress of Solid State Drives (SSDs). I am sure many of us are mesmerized by SSDs and the storage vendors are touting the best of SSDs have to offer. But let me tell you one thing – you are probably getting the least of what the best SSDs have to offer. You might be puzzled why I say things like this.

Let me share with a common sales pitch. Most (if not all) storage vendors will tout performance (usually IOPS) as the greatest benefits of SSDs. The performance numbers have to be compared to something, and that something is your regular spinning Hard Disk Drives (HDDs). The slowest SSDs in terms of IOPS is about 10-15x faster than the HDDs. A single SSD can at least churn 5,000 IOPS when compared to the fastest 15,000 RPM HDDs, which churns out about 200 IOPS (depending on HDD vendors). Therefore, the slowest SSDs can be 20-25x faster than the fastest HDDs, when measured in IOPS.

But the intent of this blogger is to share with you more about SSDs. There’s more to know because SSDs are not built the same. There are write-bias SSDs, read-bias SSDs; there are SLC (single level cell) and MLC (multi level cell) SSDs and so on. How do you differentiate them if Vendor A touts their SSDs and Vendor B touts their SSDs as well? You are not comparing SSDs and HDDs anymore. How do you know what questions to ask when they show you their performance statistics?

SNIA has recently released a set of methodology called “Solid State Storage (SSS) Performance Testing Specifications (PTS)” that helps customers evaluate and compare the SSD performance from a vendor-neutral perspective. There is also a whitepaper related to SSS PTS. This is something very important because we have to continue to educate the community about what is right and what is wrong.

In a recent webcast, the presenters from the SNIA SSS TWG (Technical Working Group) mentioned a few questions that I  think we as vendors and customers should think about when working with an SSD sales pitch. I thought I share them with you.

  • Was the performance testing done at the SSD device level or at the file system level?
  • Was the SSD pre-conditioned before the testing? If so, how?
  • Was the performance results taken at a steady state?
  • How much data was written during the testing?
  • Where was the data written to?
  • What data pattern was tested?
  • What was the test platform used to test the SSDs?
  • What hardware or software package(s) used for the testing?
  • Was the HBA bandwidth, queue depth and other parameters sufficient to test the SSDs?
  • What type of NAND Flash was used?
  • What is the target workload?
  • What was the percentage weight of the mix of Reads and Writes?
  • Are there warranty life design issue?

I thought that these questions were very relevant in understanding SSDs’ performance. And I also got to know that SSDs behave differently throughout the life stages of the device. From a performance point of view, there are 3 distinct performance life stages

  • Fresh out of the box (FOB)
  • Transition
  • Steady State

 

As you can see from the graph below, a SSD, fresh out of the box (FOB) displayed considerable performance numbers. Over a period of time (the graph shown minutes), it transitioned into a mezzanine stage of lower IOPS and finally, it normalized to the state called the Steady State. The Steady State is the desirable test range that will give the most accurate type of IOPS numbers. Therefore, it is important that your storage vendor’s performance numbers should be taken during this life stage.

Another consideration when understanding the SSDs’ performance numbers are what type of tests used? The test could be done at the file system level or at the device level. As shown in the diagram below, the test numbers could be taken from many different elements through the stack of the data path.

 

Performance for cached data would given impressive numbers but it is not accurate. File system performance will not be useful because the data travels through different layers, masking the true performance capability of the SSDs. Therefore, SNIA’s performance is based on a synthetic device level test to achieve consistency and a more accurate IOPS numbers.

There are many other factors used to determine the most relevant performance numbers. The SNIA PTS test has 4 main test suite that addresses different aspects of the SSD’s performance. They are:

  • Write Saturation test
  • Latency test
  • IOPS test
  • Throughput test

The SSS PTS would be able to reveal which is a better SSD. Here’s a sample report on latency.

Once again, it is important to know and not to take vendors’ numbers in verbatim. As the SSD market continue to grow, the responsibility lies on both side of the fence – the vendor and the customer.

 

Is there IOPS for Cloud Storage? – Nasuni style

I was in Singapore last week attending the Cloud Infrastructure Services course.

In the class, one of the foundation components of Cloud Computing is of course, storage. As the students and the instructor talked about Storage, one very interesting argument surfaced. It revolved around the storage, if it was offered on the cloud. A lot of people assumed that Cloud Storage would be for their databases, and their virtual machines, which of course, is true when the communication between the applications, virtual machines and databases are in the local area network of the Cloud Service Provider (CSP).

However, if the storage is offered through the cloud to applications that are sitting on-premise in the customer’s server room, then we have to think twice of how we perceive Cloud Storage. In this aspect, the Cloud Storage offered by the CSP is a Infrastructure-as-a-Service (IaaS), where the key service is Storage. We have to differentiate that this Storage functions as a data container, and usually not for I/O performance reasons.

Though this concept probably will be easily understood by storage professionals like us, this can cause a bit confusion for someone new to the concept of Cloud Computing and Cloud Storage. This confusion, unfortunately, is caused by many of us who are vendors or solution providers, or even publications and magazines. We are responsible to disseminate correct information to customers, but due to our lack of knowledge and experience in this extremely new market of Cloud Storage, we have created the FUDs (Fear, Uncertainty and Doubt) and hype.

Therefore, it is the duty of this blogger to clear the vapourware, and hopefully pass on the right information to accelerate  the adoption of Cloud Storage in the near future. At this moment, given the various factors such as network costs, high network latency and lack of key network technologies similar to LAN in Cloud Computing, Cloud Storage is, most of the time, for data storage containership and archiving only. And there are no IOPS or any performance related statistics related to Cloud Storage. If any engineer or vendor tells you that they have the fastest Cloud Storage in the industry, do me a favour. Give him/her a knock on the head for me!

Of course, as technologies evolve, this could change in the near future. For now, Cloud Storage is a container, NOT a high performance storage in the cloud. It is usually not meant for transactional data. There are many vendors in the Cloud Storage space from real CSPs to storage companies offering re-packaged storage boxes that are “cloud-ready”. A good example of a CSP offering Cloud Storage is Amazon S3 (Simple Storage Service). And storage vendors such as EMC and HDS are repackaging and rebranding their storage technologies as object storage, ready for the cloud. EMC Atmos is really a repackaged and rebranded Centera, with some slight modifications, while HDS , using their Archiving solution, has HCP (aka HCAP). There’s nothing wrong with what EMC and HDS have done, but before the overhyping of the world of Cloud Computing, these platforms were meant for immutable data archiving reasons. Just thought you should know.

One particular company that captured my imagination and addresses the storage performance portion is Nasuni. Of course, they are quite inventive with the Cloud Storage Gateway approach. Nasuni comes up with a Cloud Storage Gateway filer appliance, which can be either a physical 1U server or as a VMware or Hyper-V virtual appliance sitting on-premise at the customer’s site.

The key to this is “on-premise”, which allows access to data much faster because they are locally-cached in the Nasuni filer appliance itself. This Nasuni filer piece addresses the Cloud Storage “performance” piece but Nasuni do not claim any performance statistics with such implementation. The clever bit is that this addresses data or files that are transactional in nature, i.e. NFS or CIFS, to serve data or files “locally”. (I wonder if Nasuni filer has iSCSI as well. Hmmmm….)

In the Nasuni architecture, they “break up” their “Cloud Storage” into 2 pieces. Piece #1 sits on-premise, at the customer site, and acts as a bridge to the Piece #2, that is sitting in a Cloud Storage. From a simplified view, have a look at the diagram below:

 

 

Piece #1 is the component that handles some of the transactional traffic related to files. In a more technical diagram below, you can see that the Nasuni filer addresses the file sharing portion, using the local disks on the filer appliance as a local caching mechanism.

 

Furthermore, older file pieces are whiffed away to the any Cloud Storage using the Cloud Connector interface, hence giving the customer a sense that their storage capacity needs can be limitless if they want to (for a fee, of course). At the same time, the Nasuni filer support thin provisioning and snapshots. How cool is that!

The Cloud Storage piece (Piece #2) is used for the data container and archiving reasons. This component can be sitting and hosted at Amazon S3, Microsoft Azure, Rackspace Cloud Files, Nirvanix Storage Delivery Network and Iron Mountain Archive Services Platform.

The data communication and transfer between the Nasuni filer is secure, encrypted, deduplication and compressed, giving it the efficiency and security that most customers would be concerned about. The diagram below explains the dat communication and data transfer bit.

 

In this manner, the Nasuni filer can replace traditional NAS platforms and can potentially provide a much lower total cost of ownership (TCO) in the long run. Nasuni does not pretend to be a NAS replacement. To me, this concept is very inventive and could potentially change the way we perceive file sharing and file server, obscuring and blurring concept of NAS.

Again, I would like to reiterate that Nasuni does not attempt to say their solution is a NAS or a performance-based Cloud Storage but what they have cleverly packaged seems to be appealing to customers. Their customer base has grown 78% in Q2 of 2011. It’s just too bad they are not here in Malaysia or this part of the world (yet).

IOPS in Cloud Storage? Not yet.

 

The recipe for storage performance modeling

Good morning, afternoon, evening, Ladies & Gentlemen, wherever you are.

Today, we are going to learn how to bake, errr … I mean, make a storage performance model. Before we begin, allow me to set the stage.

Don’t you just hate it when you are asked to do storage performance sizing and you don’t have a freaking idea how to get started? A typical techie would probably say, “Aiya, just use the capacity lah!”, and usually, they will proceed to size the storage according to capacity. In fact, sizing by capacity is the worst way to do storage performance modeling.

Bear in mind that storage is not a black box, although some people wished it was. It is not black magic when it comes to performance sizing because things can be applied in a very scientific and logical manner.

SNIA (Storage Networking Industry Association) has made a storage performance modeling methodology (that’s quite a mouthful), and basically simplified it into these few key ingredients. This recipe is for storage performance modeling in general and I am advising you guys out there to engage your storage vendors professional services. They will know their storage solutions best.

And I am going to say to you – Don’t be cheap and not engage professional services – to get to the experts out there. I was having a chat with an consultant just now at McDonald’s. I have known this friend of mine for about 6-7 years now and his name is Sugen Sumoo, the Director of DBORA Consulting. They specialize in Oracle and database performance tuning and performance forecasting and it is something that a typical DBA can’t do, because DBORA Consulting is the Professional Service that brings expertise and value to Oracle customers. Likewise, you have to engage your respective storage professional services as well.

In a cook book or a cooking show, you are presented with the ingredients used and in this recipe for storage performance modeling, the ingredients (in no particular order) are:

  • Application block size
  • Read and Write ratio
  • Application access patterns
  • Working set size
  • IOPS or throughput
  • Demand intensity

Application Block Size

First of all, the storage is there to serve applications. We always have to look from the applications’ point of view, not storage’s point of view.  Different applications have different block size. Databases typically range from 8K-64K and backup applications usually deal with larger block sizes. Video applications can have 256K block sizes or higher. It all depends.

The best way is to find out from the DBA, email administrator or application developers. The unfortunate thing is most so-called technical people or administrators in Malaysia doesn’t have a clue about the applications they manage. So, my advice to you storage professionals, do your research on the application and take the default value. These clueless fellas are likely to take the default.

Read and Write ratio

Applications behave differently at different times of the day, and at different times of the month (no, it’s not PMS). At the end of the financial year or calendar, there are some tasks that these applications do as well. But in a typical day, there are different weightage or percentage of read operations versus write operations.

Most OLTP (online transaction processing)-based applications tend to be read heavy and write light, but we need to find out the ratio. Typically, it can be a 2:1 ratio or 60%:40%, but it is best to speak to the application administrators about the ratio. DSS (Decision Support Systems) and data warehousing applications could have much higher reads than writes while a seismic-analysis applications can have multiple writes during the analysis periods. It all depends.

To counter the “clueless” administrators, ask lots of questions. Find out the workflow of several key tasks and ask what that particular tasks do at different checkpoints of the application’s processing. If you are lazy (please don’t be lazy, because it degrades your value as a storage professional), use a rule of thumb.

Application access patterns

Applications behave differently in general. They can be sequential, like backup or video streaming. They can be random like emails, databases at certain times of the day, and so on. All these behavioral patterns affect how we design and size the disks in the storage.

Some RAID levels tend to work well with sequential access and others, with random access. It is not difficult to find out about the applications’ pattern and if you read more about the different RAID-levels in storage, you can easily identify the type of RAID levels suitable for each type of behavioral patterns.

Working set size

This variable is a bit more difficult to determine. This means that a chunk of the application has to be loaded into a working area, usually memory and cache memory, to be used and abused by the application users.

Unless someone is well versed with the applications, one would not be able to determine how much of the applications would be placed in memory and in cache memory. Typically, this can only be determined after the application has been running for some time.

The flexibility of having SSDs, especially the DRAM-type of SSDs, are very useful to ensure that there is sufficient “working space” for these applications.

IOPS or Throughput

According to SNIA model, for I/O less than 64K, IOPS should be used as a yardstick to do storage performance modeling. Anything larger, use throughput, in which MB/sec is the measurement unit.

The application guy would be able to tell you what kind of IOPS their application is expecting or what kind of throughput they want. Again, ask a lot of questions, because this will help you determine the type of disks and the kind of performance you give to the application guys.

If the application guy is clueless again, ask someone more senior or ask the vendor. If the vendor engineers cannot give you an answer, then they should not be working for the vendor.

Demand intensity

This part is usually overlooked when it comes to performance sizing. Demand intensity refers to how intense is the I/O requests. It could come from 1 channel or 1 part of the applications, or it could come from several parts of the applications in parallel. It is as if the storage is being ‘bombarded’ by applications and this is the part that is hard to determine as well.

In some applications, the degree of intensity or parallelism can be tuned and to find out, ask the application administrator or developer. If not, ask the vendor. Also do a lot of research on the application’s architecture.

And one last thing. What I have learned is to add buffers to the storage performance model. Typically I would add about 10-20% extra but you never know. As storage professionals, I would strongly encourage to engage professional services, because it is worthwhile, especially in the early stages of the sizing. It is usually a more expensive affair to size it after the applications have been installed and running.

“Failure to plan is planning to fail”.  The recipe isn’t that difficult. Go figure it out.

NFS deserves more credit from guys doing virtualization

I was at the RedHat Forum last week when I chanced upon a conversation between an attendee and one of the ECS engineers. The conversation went like this

Attendee: Is the RHEV running on SAN or NAS?

ECS Engineer: Oh, for this demo, it is running NFS but in production, you should run iSCSI or Fibre Channel. NFS is only for labs only, not good for production.

Attendee: I see … (and he went off)

I was standing next to them munching my mini-pizza and in my mind, “Oh, come on, NFS is better than that!”

NAS has always played a smaller brother to SAN but usually for the wrong reasons. Perhaps it is the perception that NAS is low-end and not good enough for high-end production systems. However, this is very wrong because NAS has been growing at a faster rate than Fibre Channel, and at the same time Fibre Channel growth has been tapering and possibly on the wane. And I have always said that NAS is a better suited protocol when it comes to unstructured data and files because the NAS protocol is the new storage networking currency of Internet storage and the Cloud (this could change very soon with the REST protocol, but that’s another story). Where else can you find a protocol where sharing is key. iSCSI, even though it has been growing at a faster pace in production storage, cannot be shared easily because it is block-based.

Now back to NFS. NFS version 3 has been around for more than 15 years and has taken its share of bad raps. I agree that this protocol is still very much in the landscape of most NFS installations. But NFS version 4 is changing all that taking on the better parts of the CIFS protocol, notably the equivalent of opportunistic locking or oplocks. In addition to that it has greatly enhanced its security, incorporating Kerberos-type of authentication. As for performance, NFS v4 added in a compounded in a COMPOUND operations for aggregating operations into a single request.

Today, most virtualization solutions from VMware and RedHat works with NFS natively. Note that the Windows CIFS protocol is not supported, only NFS.

This blog entry is not stating that NFS is better than iSCSI or FC but to give NFS credit where credit is due. NFS is not inferior to these block-based protocols. In fact, there are situations where NFS is better, like for instance, expanding the NFS-based datastore on the fly in a VMware implementation. I will use several performance related examples since performance is often used as a yardstick when these protocols are compared.

In an experiment conducted by VMware based on a version 4.0, with all things being equal, below is a series of graphs that compares these 3 protocols (NFS, iSCSI and FC). Note the comparison between NFS and iSCSI rather than FC because NFS and iSCSI run on Gigabit Ethernet, whereas FC is on a different networking platform (hey, if you got the money, go ahead and buy FC!)

Based a one virtual machine (VM), the Read throughput statistics (higher is better) are:

 

The red circle shows that NFS is up there with iSCSI in terms of read throughput from 4K blocks to 512K blocks. As for write throughput for 1 VM, the graph is shown below:


Even though NFS suffers in write throughput in the smaller blocks less than 16KB, NFS performance write throughput improves over iSCSI when between 16K and 32K range and is equal when it is in 64K, 128K and 512K block tests.

The 2 graphs above are of a single VM. But in most real production environment, a single ESX host will run multiple VMs and here is the throughput graph for multiple VMs.

Again, you can see that in a multiple VMs environment, NFS and iSCSI are equal in throughput, dispelling the notion that NFS is not as good in performance as iSCSI.

Oh, you might say that this is just VMs without any OSes or any applications running in these VMs. Next, I want to share with you another performance testing conducted by VMware for an Microsoft Exchange environment.

The next statistics are produced from an Exchange Load Generator (popularly known as LoadGen) to simulate the load of 16,000 Exchange users running in 8 VMs. With all things being equal again, you will be surprised after you see these graphs.

The graph above shows the average send mail latency of the 3 protocols (lower is better). On the average, NFS has lower latency than iSCSI, better than what most people might think. Another graph shows the 95th percentile of send mail latency below:

 

Again, you can see that the NFS’s latency is lower than iSCSI. Interesting isn’t it?

What about IOPS then? In another test with an 8-hour DoubleHeavy LoadGen simulator, the IOPS graphs for all 3 protocols are shown below:

In the graph above (higher is better), NFS performed reasonably well compared to the other 2 block-based protocols, and even outperforming iSCSI in this 8-hour load testing. Surprising huh?

As I have shown, NFS is not inferior compared to the block-based protocols such as iSCSI. In fact, VMware in version 4.1 has improved all 3 storage protocols significantly as mentioned in the VMware paper. The following are quoted in the paper for NFS and iSCSI.

  1. Using storage microbenchmarks, we observe that vSphere 4.1 NFS shows improvements in the range of 12–40% for Reads,and improvements in the range of 32–124% for Writes, over 10GbE.
  2. Using storage microbenchmarks, we observe that vSphere 4.1 Software iSCSI shows improvements in the range of 6–23% for Reads, and improvements in the range of 8–19% for Writes, over 10GbE

The performance improvement for NFS is significant when the network infrastructure was 10GbE. The percentage jump between 32-124%! That’s a whopping figure compared to iSCSI which ranged from 8-19%. Since both protocols are neck-to-neck in version 4.0, NFS seems to be taking a bigger lead in version 4.1. With the release of VMware version 5.0 a few weeks ago, we shall know the performance of both NFS and iSCSI soon.

To be fair, NFS does take a higher CPU performance hit compared to iSCSI as the graph below shows:

Also note that the load testing are based on NFS version 3. If version 4 was used, I am sure the performance statistics above will take a whole new plateau.

Therefore, NFS isn’t inferior at all compared to iSCSI, even in a 10GbE environment. We just got to know the facts instead of brushing off NFS.

What kind of IOPS and throughput do you get from RAID-5/6? – Part 2

In my previous blog entry, I mentioned the write penalty for RAID-5/6. This factor will figure heavily in the way we size the RAID-level for performance capacity planning.

It is difficult to ascertain what kind of IOPS and throughput that are required for an application, especially a database, to run well with additional room to grow. From a DBA or an application developer, I believe they would have adequate information to tell what is the numbers of users that the application can support, both average and peak, transactions per second (TPS), block size required for logs, database files and so on.

But as we are all aware, most of the time, these types of information are not readily available. So, coming from a storage angle, the storage administrator can advise the DBA or the application developer that the configured RAID group or volume or LUN is capable of delivering a certain number of IOPS and is able to achieve a certain throughput MB/sec. These numbers will be off the box itself immediately. Of course, other factors such as HBA speed, the FC/iSCSI configurations, the network traffic and so on will affect the overall performance delivery to the application. But we can safely inform the DBA and/or the application developer that this is what the storage is delivering out of the box.

The building blocks of all storage RAID groups/volumes/LUNs are pretty much your hard disk drives (HDDs) and/or Solid State Drives (SSDs). The manufacturer of these disks will usually publish the IOPS and throughput of individual drives but if these information is not available, we can construct IOPS of an individual HDD from its seek and latency times.

For example, if the HDD’s

average latency = 2.8 ms;          average read seek = 4.2 ms;              average write seek = 4.8 ms

then the IOPS can be calculated as

                                  1
         IOPS = ---------------------------------------
                (average latency) + (average seek time)

Therefore from the details above,

                    1
         IOPS = -------------------  = 136.986 IOPS
                (0.0028) + (0.0045)

That’s pretty simple, right? But of course, it is easier to just accept that a certain type of disk will have a range of IOPS as shown in the table below:

Disk Type RPM IOPS Range
SATA 5,400 50-75
SATA 7,200 75-100
SAS/FC 10,000 100-125
SAS/FC 15,000 175-200
SSD N/A 5,000-10,000

The information from the table above is just for reference only and by no means a very accurate one but it is good enough for us to determine the IOPS of a RAID group/volume/LUN. Let’s look at the RAID write penalty again in the table below:

RAID-level Number of I/O Reads
Number of I/O for Writes
RAID Write Penalty
0 1 1 1
1 (1+0, 0+1) 1 2 2
5 1 4 4
6 1 6 6

Next, we need to know what is the ratio of Reads vs Writes for that particular database or application. I mentioned earlier that in OLTP-type of applications, we usually take a 2:1 or 3:1 ratio in favour of Reads.

To make things simpler, let’s assume we create a RAID-6 volume of 6 data disks and 2 parity disks in a RAID-6 (6+2) configuration. The disks used are SATA disks of 7,200 RPM, with each individual disk of 100 IOPS. Assume we are using a ratio of 2:1 in favour of Reads, which gives us 66.666% and 33.333% respectively for Reads and Writes.

Therefore, the combined IOPS of the 8 disks in the RAID-6 configuration is probably about 800 IOPS. However, because of the write penalty of RAID-6, the effective IOPS for the RAID-6 volume will be lower than that. Let’s do some calculation to see what happens:

1)  Read IOPS + Write IOPS = 800 IOPS

2)  (0.66666 x 800) + (0.33333 x 800) = 800 IOPS

3) Read IOPS will be 0.66666 x 800 = 533.328 IOPS

4) Write IOPS will be 0.33333 x 800 = 266.664 IOPS. However, since RAID-6 has a write penalty of 6, this number has to be divided by 6. 266.664/6 will be 44.444 IOPS for Writes

Therefore, what the RAID-6 volume is capable of is approximately 533 IOPS for Reads and 44 IOPS for Writes.

We have determined IOPS for the RAID volume but what about throughput. Throughput is determined by the block size used. Assume that our RAID-6 volume uses a 4-K block size. With a combined effective IOPS of 577 (533+44), we multiply the IOPS with the block size

     Throughput = 577 IOPS x 4-KB
                = 2308KB/sec

Therefore when I/O is sustained in a sequential manner, the effective throughput is 2308KB/sec.

On the other hand, we often were told to add more spindles to the volume to increase the IOPS. This is true, to a point, where the maximum amount of IOPS that can be delivered will taper into a flatline, because the I/O channel to the RAID volume  has been saturated. Therefore, it is best to know that adding more spindles does not always equate to a higher IOPS.

Performance sizing for a database or an application is both a science and an art. Mathematically, we can prove things to a a certain amount of accuracy and confidence but each storage platform is very different in the way they handle RAID. Newer storage platforms have proprietary RAID that nowadays, it does not matter much what kind of RAID is best for the application. Vendors such as IBM XIV has RAID-X which both radical in design and implementation. NetApp will almost always say RAID-DP is the best no matter what, because RAID-DP is all NetApp.

So there is no right or wrong to choose the RAID-level for the application. But it is VERY important to know what are the best practice are and my advice is everyone is to do Proof-of-Concepts, and TEST, TEST, TEST! And ASK QUESTIONS!

Don’t RAID-5/6 everything! – Part 1

It’s a beautiful Saturday morning … the sun is out, and the birds are chirping … and here I am, thinking about RAID-5/6. What’s wrong with me?

Anyway, have you ever wondered almost all your volumes are in a RAID-5/6 configuration? Like an obedient child, the answer would probably be “Oh, my vendor said it is good for me …”

In storage, the rule is applications-read, applications-write. And different applications have different behaviors but typically, they fall under 2 categories:

  • Random access
  • Sequential access

The next question to ask is how much Read/Writes ratio (or percentage) is in that Random Access behavior and how much of Read/Write ratio in Sequential Access behavior.

We usually pigeonhole transactional databases such as SQL Server, Oracle into OLTP-type characteristics with random access being the dominant access method. Similarly, email applications such as Exchange, Lotus and even SMTP into similar OLTP-type characteristics as well. We typically do a 2:1 or 3:1 ratio for OLTP-type applications with Read heavy and less of Writes. Data warehouse type of databases tend to be more sequential.

However, even within these OLTP applications, there are also sequential access behaviors as well, as the following table for a database shows:

Operation Random or Sequential Read/Write Heavy Block Size
DB-Log Random (Sequential in log recovery) Write Heavy unless you are doing log recovery 1KB – 64KB
DB-Data Files Random Read/Write mix dependent on load 4KB – 32KB
Batch insert Sequential Write Heavy 8KB – 128KB
Index scan Sequential Read Heavy 8KB – 128KB

We will look into 4 RAID-levels in this scenario and see how each RAID-level applies to an OLTP-type of environment. These RAID levels are RAID-0, RAID-1 (1+0, 0+1 included), RAID-5 and RAID-6.

RAID-0 is the baseline, with 1 x Read and 1 x Write being processed as per normal.

In RAID-1, it would require 2 x Writes and 1 x Read, because the write operation is mirrored. The RAID penalty is 2.

To avoid the cost of RAID-1, RAID-5 is almost always the RAID level of choice (unless you speak to those NetApp fellas). RAID-5 is a parity-based RAID and require 2 x Read (1 to read the data block and 1 to read the parity block) AND 2 x Write (1 to write the modified block and 1 to write the modified parity). Hence it has a RAID penalty of 4.

RAID-6 was to address the risk of RAID-5 because disk capacity are so freaking large now (3TB just came out). To rebuild a large-TB drive would take longer time and the RAID-5 volume is at risk if a second disk failure occurs. Hence, double parity RAID in RAID-6. But unfortunately, the RAID penalty for RAID-6 is 6!

To summarize the RAID write penalty,

RAID-level Number of I/O Reads
Number of I/O for Writes
RAID Write Penalty
0 1 1 1
1 (1+0, 0+1) 1 2 2
5 1 4 4
6 1 6 6

So, it is well known that RAID 0 has good performance for reads and writes but with absolutely no protection. RAID-1 would be good for random reads and writes but it is costly. RAID-5 is good for applications with a high ratio of sequential reads vs writes (2:1, 3:1 as mentioned), and RAID-6, errr … should be taken similarly as RAID-5 with some additional performance penalty.

With that in mind, a storage administrator must question why a particular RAID-level was proposed to the database or any like-applications.

I am going out to enjoy the Saturday now … and today, August 13th is the World’s Left-Handed Day. More about this RAID penalty and IOPS in my next entry.