Tag Archives: performance

Networking, Storage

Ferraris and Tractor Trailers

Leave a comment

Working in the SMB world, it is actually pretty rare that we need to talk about latency.  The SMB world is almost universally focused on system throughput and generally unaware of latency as a need.  But there are times where latency becomes important and when it does it is critical that we understand the interplay of throughput and latency and just what “speed” means to us.  Once we start moving into the enterprise space, latency is more often going to be viewed as a concern, but even there throughput nearly always reigns supreme, to the point that concepts of speed almost universally revolve around throughput and concepts of latency are often ignored or forgotten.

Understanding the role of latency in a system can be complicated, even though latency itself is relatively simple to understand.

A great comparison between latency and throughput that I like to use is the idea of a Ferrari and a tractor trailer.  Ferraris are “fast” in the traditional sense, they have a high “miles per hour.”  One might say that they are designed for speed.  But are they?

We generally consider tractor trailers to be slow.  They are big and lumbering beasts that have a low top end speed.  But they haul a lot of stuff at once.

In computer terms we normally think of speed like hauling capacity – we think in terms of “items” per second.  In the terms of a Ferrari going two hundred miles per hour is great, but it can haul maybe one box at a time.  A tractor trailer can only go one hundred miles per hour but can haul closer to one thousand boxes at a time.  When we talk about throughput or speed on a computer this is more what we think about.  In network terms we think of gigabytes per second and are rarely concerned with the speed of an individual packet as a single packet is rarely important.  In computational terms we think about ideas like floating point operations per second, a similar concept.  No one really cares how long a single FLOP (floating point operation) takes, only how many we can get done in one or ten seconds.

So when looking at a Ferrari we could say that it has a useful speed of two hundred box-miles per hour.  That is for every hour of operations, a Ferrari can move one box up to two hundred miles.  A tractor trailer has a useful speed of one hundred thousand box-miles per hour.  In terms of moving packages around, the throughput of the tractor trailer is easily five hundred times “faster” than that of the Ferrari.

So in terms of how we normally think of computers and networks a tractor trailer would be “fast” and a Ferrari would be “slow.”

But there is also latency to consider.  Assuming that our payload is tiny, say a letter or a small box, a Ferrari can move that one box over a thousand miles in just five hours!  A tractor trailer would take ten hours to make this same journey (but could have a LOT of letters all arriving at once.)  If what we need is to get a message or a small parcel from one place to another very quickly the Ferrari is the better choice because it has half the latency (delay) from the time we initiate the delivery until the first package is delivered than the tractor trailer does.

As you can imagine, in most cases tractor trailers are vastly more practical because their delivery speed is so much higher.  And, this being the case, we actually see large trucks on the highways all of the time and the occurrence rate of Ferraris is very low – even though each cost about the same amount to purchase (very roughly.)  But in special cases, the Ferrari makes more sense.  Just not very often.

This is a general case concept and can apply to numerous applications.  It applies to caching systems, memory, CPU, networking, operating system kernels and schedulers, to cars and more.  Latency and throughput are generally inversely related – we give up latency in order to obtain throughput.  For most operations this makes the best sense.  But sometimes it makes more sense to tune for latency.

Storage is actually an odd duck in computing where nearly all focus on storage performance is around IOPS, which is roughly a proxy measurement for latency, instead of throughput which is measured in “data transferred per second.”  Rarely do we care about this second number as it is almost never the source of storage bottlenecks.  But this is the exception, not the rule.

Latency and throughput can have some surprising interactions in the computing world.  When we talk about networks, for example, we typically measure only throughput (Gb/s) but rarely care much about the latency (normally measured in milliseconds.)  Typically this is because nearly all networking systems have similar latency numbers and most applications are pretty much unconcerned with latency delays.  It is only the rare application like VoIP over International links or satellite where latency affects the average person or can sometimes surprise people when they attempt something uncommon like iSCSI over a long distance WAN connection and suddenly latency pops up to surprise them as an unforeseen problem.

One of the places where the interaction of latency and throughput starts to become shocking and interesting is when we move from electrical or optical data networks to physical ones.  A famous quote in the industry is:

Never underestimate the bandwidth of a station wagon full of tapes hurtling down the highway.

This is a great demonstration of huge bandwidth with very high latency.  Driving fifty miles across town a single stationwagon or SUV could haul hundreds of petabytes of data hitting data rates that 10GB/s fiber could not come close to.  But the time for the first data packet to arrive is about an hour.  We often discount this kind of network because we assume that latency must be bounded at under about 500ms.  But that is not always the case.

Australia recently made the news where they did a test to see if a pigeon carrying an SD card could, in terms of network throughput, outperform the regions ISP – and the pigeon ended up being faster than the ISP!

In terms of computing performance we often ignore latency to the point of not even being aware of it as a context in which to discuss performance.  But in low latency computing circles it is considered very carefully.  System throughput is generally greatly reduced (it becomes common to target systems to only hit ten percent CPU utilization when more traditional systems target closer to ninety percent) with concepts like real time kernels, CPU affinity, processor pinning, cache hit ratios and lowered measuring all being used to focus on obtaining the most immediate response possible from a system rather than attempting to get the most total processing out of a system.

Common places where low latency from a computational perspective is desired is in critical controller systems (such as manufacturing controllers were even a millisecond of latency can cause problems on the factory floor) or in financial trading systems where a few milliseconds of delay can cause investments to have changed in price or products to have already been sold and no longer be available.  Speed, in terms of latency, is often the deciding factor between making money or losing money – even a single millisecond can be crippling.

Technically even audio and video processing systems have to be latency sensitive but most modern computing systems have so much spare processing overhead and latency is generally low enough that most systems, even VoIP PBXs and conferencing systems, can function today with only very rarely needing to be aware of latency concerns on the processing side (even networking latency is becoming less and less common as a concern.)  The average system administrator or engineer might easily manage to go through a career without ever needing to work on a system that is latency sensitive or for which there is not so much available overhead as to hide any latency sensitivity.

Defining speed, whether that means throughput, latency or even something else or some combination of the two, is something that is very important in all aspects of IT and in life.  Understanding how they affect us in different situations and how they react to each other with them generally existing in an indirect relationship where improvements in throughput come at a cost to latency or vice versa and learning to balance these as needed to improve the systems that we work on is very valuable.

Storage

Practical RAID Performance

1 Comment

Choosing a RAID level is an exercise in balancing many factors including cost, reliability, capacity and, of course, performance.  RAID performance can be difficult to understand especially as different RAID levels use different techniques and behave rather differently from each other in some cases.  In this article I want to explore the common RAID levels of RAID 0, 5, 6 and 10 to see how performance differs between them.

For the purposes of this article, RAID 1 will be assumed to be a subset of RAID 10.   This is often a handy way to think of RAID 1 – as simply being a RAID 10 array with only a single mirrored pair member.  As RAID 1 is truly a single pair RAID 10 and behaves as such this works wonderfully for making RAID performance easy to understand as it simply maps into the RAID 10 performance curve.

There are two types of performance to look at with all storage: reading and writing.  In terms of RAID reading is extremely easy and writing is rather complex.  Read performance is effectively stable across all RAID types.  Writing, however, is not.

To make discussing performance easier we need to define a few terms as we will be working with some equations. In our discussions we will use N to represent the total number of drives, often referred to as spindles, in our array and we will use X to refer to the performance of each drive individually.  This allows us to talk in terms of relative performance as a factor of the drive performance allowing us to abstract away the RAID array and not have to think in terms of raw IOPS.  This is important as IOPS are often very hard to define but we can compare performance in a meaningful way by speaking to it in relationship to the individual drives within the array.

It is also important to remember that we are only talking about the performance of the RAID array itself, not an entire storage subsystem.  Artifacts such as memory caches and solid state caches will do amazing things to alter the overall performance of a storage subsystem, but do not fundamentally change the performance of the RAID array itself under the hood.  There is no simple formula for determining how different cache options will impact the overall performance but suffice it to say that it can be very dramatic but this depends heavily not only on the cache choices themselves but also heavily upon workload. Even the biggest, fastest, most robust cache options cannot change the long term, sustained performance of an array.

RAID is complex and many factors influence the final performance.  One is the implementation of the RAID system itself.  A poor implementation might cause latency or may fail to take advantage of the available spindles (such as having a RAID 1 array read only from a single disk instead of from both simultaneously!)  There is no easy way to account for deficiencies in specific RAID implementations so we must assume that all are working to the limits of the specification as, indeed, any enterprise RAID system will do. It is primarily hobby and consumer RAID systems that fail to do this.

Some types of RAID also have dramatic amounts of computational overhead associated with them while others do not.  Primarily parity RAID levels require heavy processing in order to handle write operations with different levels having different amounts of computation necessary for each operation.  This introduces latency, but does not curtail throughput.  This latency will vary, however, based on the implementation of the RAID level as well as on the processing capability of the system in question.  Hardware RAID will use something like a general purpose CPU (often a Power or ARM RISC processor) or a custom ASIC to handle this while software RAID hands this off to the server’s own CPU.  Often the server CPU is actually faster here but consumes system resources.  ASICs can be very fast but are expensive to produce.  This latency impacts storage performance but is very difficult to predict and can vary from nominal to dramatic.  So I will mention the relative latency impact with each RAID level but will not attempt to measure it.  In most RAID performance calculations, this latency is ignored but it is important to understand that it is present and could, depending on the configuration of the array, have a noticeable impact on a workload.

There is, it should be mentioned, a tiny performance impact to read operations due to efficiencies in the layout of data on the disk itself.  Parity RAID requires there to be data on the disks that is useless during a healthy read operation but cannot be used to speed it up.  The actually results in it being slightly slower.  But this impact is ridiculously small and is normally not measured and so can be ignored.

Factors such as stripe size also impact performance, of course, but as that is configurable and not an intrinsic artifact in any RAID level I will ignore it here.  It is not a factor when choosing a RAID level itself but only in configuring one once chosen.

The final factor that I want to mention is the read to write ratio of storage operations.  Some RAID arrays will be used almost purely for read operations, some almost solely for write operations but most use a blend of the two, likely something like eighty percent read and twenty percent write.  This ratio is very important in understanding the performance that you will get from your specific RAID array and understanding how each RAID level will impact you.  I refer to this as the read/write blend.

We measure storage performance primarily in IOPS.  IOPS stands for Input/Output Operations Per Second (yes, I know that the letters don’t line up well, it is what it is.)  I further use the terms RIOPS for Read IOPS, WIOPS for Write IOPS and BIOPS for Blended IOPS which would come with a ration 80/20 or whatever.  Many people talk about storage performance with a single IOPS number.  When this is done they normally mean Blended IOPS at 50/50.  However, rarely does any workload run at 50/50 so that number can be extremely misleading.  Two numbers, RIOPS and WIOPS is what is needed to understand performance and these two together can be used to find any IOPS Blend that is needed.   For example, a 50/50 blend is as simple as (RIOPS * .5) + (WIOPS * .5).  The more common 80/20 blend would be (RIOPS * .8) + (WIOPS * .2).

Now that we have established some criteria and background understanding we will delve into our RAID levels themselves and see how performance varies across them.

For all RAID levels, the Read IOPS number is calculated using NX.  This does not address the nominal overhead numbers that I mention above, of course.  This is a “best case” number but the real world number is so close that it is very practical to simply use this formula.  Since take the number of spindles (N) and multiple by the IOPS performance of an individual drive (X).  Keep in mind that drives often have different read and write performance so be sure to use the drives Read IOPS rating or tested speed for the Read IOPS calculation and the Write IOPS rate or tested speed for the Write IOPS calculation.

RAID 0

RAID 0 is the easiest RAID level to understand because there is effectively no overhead to worry about, no resources consumed to power it and both read and write get the full benefit of every spindle, all of the time.  So for RAID 0 our formula for write performance is very simple: NX.  RAID 0 is always the most performant RAID level.

An example would be an eight spindle RAID 0 array.  If an individual drive in the array delivers 125 IOPS then our calculation would be from N = 8 and X = 125 so 8 * 125 yielding 1,000 IOPS.  Since both read and write IOPS are the same here, it is extremely simple as we get 1K RIOPS, 1K WIOPS and 1K with any blending thereof.  Very simple.  If we didn’t know the absolute IOPS of an individual spindle we could refer to an eight spindle RAID 0 as delivering 8X Blended IOPS.

RAID 10

RAID 10 has the second simplest RAID level to calculate.  Because RAID 10 is a RAID 0 stripe of mirror sets, we have no overhead to worry about from the stripe but each mirror has to write the same data twice in order to create the mirroring.  This cuts our write performance in half compared to a RAID 0 array of the same number of drives.  Giving us a write performance formula of simply: NX/2  or .5NX.

It should be noted that at the same capacity, rather than the same number of spindles, RAID 10 has the same write performance as RAID 0 but double the read performance – simply because it requires twice as many spindles to match the same capacity.

So an eight spindle RAID 10 array would be N = 8 and X = 125 and our resulting calculation comes out to be (8 * 125)/2 which is 500 WIOPS or 4X WIOPS.  A 50/50 blend would result in 750 Blended IOPS (1,000 Read IOPS and 500 Write IOPS.)

This formula applies to RAID 1, RAID 10, RAID 100 and RAID 01 equally.

Uncommon options such as triple mirroring in RAID 10 would alter this write penalty.  RAID 10 with triple mirroring would be NX/3, for example.

RAID 5

While RAID 5 is deprecated and should never be used in new arrays I include it here because it is a well known and commonly used RAID level and its performance needs to be understood.  RAID 5 is the most basic of the modern parity RAID levels.  RAID 2, 3 & 4 are no longer found in production systems and so we will not look into their performance here.  RAID 5, while not recommended for use today, is the foundation of other modern parity RAID levels so is important to understand.

Parity RAID adds a somewhat complicated need to verify and re-write parity with every write that goes to disk.  This means that a RAID 5 array will have to read the data, read the parity, write the data and finally write the parity.  Four operations for each effective one.  This gives us a write penalty on RAID 5 of four.  So the formula for RAID 5 write performance is NX/4.

So following the eight spindle example where the write IOPS of an individual spindle is 125 we would get the following calculation: (8 * 125)/4 or 2X Write IOPS which comes to 250 WIOPS.  In a 50/50 blend this would result in 625 Blended IOPS.

RAID 6

RAID 6, after RAID 10, is probably the most common and useful RAID level in use today.  RAID 6, however, is based off of RAID 5 and has another level of parity.  This makes it dramatically safer than RAID 5, which is very important, but also imposes a dramatic write penalty as each write operation requires the disks to read the data, read the first parity, read the second parity, write the data, write the first parity and then finally write the second parity.  This comes out to be a six times write penalty, which is pretty dramatic.  So our formula is NX/6.

Continuing our example we get (8 * 125)/6 which comes out to ~167 Write IOPS or 1.33X.  In our 50/50 blend example this is a performance of  583.5 Blended IOPS.  As you can see, parity writes cause a very rapid decrease in write performance and a noticeable drop in blended performance.

RAID 7 (aka RAID 5.3 or RAID 7.3)

RAID 7 is a somewhat non-standard RAID level with triple parity based off of the existing single parity of RAID 5 and the existing double parity of RAID 6.  The only current implementation of RAID 7 is ZFS’ RAIDZ3.  Because RAID 7 contains all of the overhead of both RAID 5 and RAID 6 plus the additional overhead of the third parity component we have a write penalty of a staggering eight times.  So our formula for finding RAID 7 write performance is NX/8.

In our example this would mean that (8 * 125)/8 would come out to 125 Write IOPS or 1X.  So with eight drives in our array we would get only the write performance of a single, stand alone drive.  That is significant overhead.  Our blended 50/50 IOPS would come out to only 562.5.

Complex RAID

Complex RAID levels or Nested RAID levels such as RAID 50, 60, 61, 16, etc. can be found using the information above and breaking the RAID down into its components and applying each using the formulæ provided above.  There is no simple formula for these levels because they have varying configurations.  It is necessary to break them down into their components and apply the formulæ multiple times.

RAID 60 with twelve drives, two sets of six drives, where each drive is 150 IOPS would be done with two RAID 6s.  It would be the NX of RAID 0 where N is two (for two RAID 6 arrays) and the X is the resultant performance of each RAID 6.   Each RAID 6 set would be (6 * 150)/6.  So the full array would be 2((6 * 150)/6).  Which results in 300 Write IOPS.

The same example as above but configured as RAID 61, a mirrored pair of RAID 6 arrays, would be the same performance per RAID 6 array, but applied to the RAID 1 formula which is NX/2 (where X is the resultant performance of the each RAID array.)  So the final formula would be 2((6 * 150)/6)/2 coming to 150 Write IOPS from twelve drives.

Performance as a Factor of Capacity

When we are producing RAID performance formulæ we think of these in terms of the number of spindles which is incredibly sensible.  This is very useful in determining the performance of a proposed array or even an existing one where measurement is not possible and allows us to compare the relative performance between different proposed options.  It is in these terms that we universally think of RAID performance.

This is not always a good approach, however, because typically we look at RAID as a factor of capacity rather than of performance or spindle count.  It would be very rare, but certainly possible, that someone would consider an eight drive RAID 6 array versus an eight drive RAID 10 array.  Once in a while this will occur due to a chassis limitation or some other, similar reason.  But typically RAID arrays are viewed from the standpoint of total array capacity (e.g. usable capacity) rather than spindle count, performance or any other factor.  It is odd, therefore, that we should then switch to viewing RAID performance as a function of spindle count.

If we change our viewpoint and pivot upon capacity as the common factor, while still assuming that individual drive capacity and performance (X) remains constant between comparators then we arrive at a completely different landscape of performance.  In doing this we see, for example, that RAID 0 is no longer the most performant RAID level and that read performance varies dramatically instead of being a constant.

Capacity is a fickle thing but we can distill it out to the number of spindles necessary to reach desired capacity.  This makes this discussion far easier.  So our first step is to determine our spindle count needed for raw capacity.  If we need a capacity of 10TB and are using 1TB drives, we would need ten spindles, for example.  Or if we need 3.2TB and are using 600GB drives we would need six spindles.  We will, different than before, refer to our spindle count as R.  As before, performance of the individual drive is represented as X.  (R is used here to denote that this is the Raw Capacity Count, rather that the total Number of spindles.)

RAID 0 remains simple, performance is still RX as there are no additional drives.  Both read and write IOPS are simply NX.

RAID 10 has RX Write IOPS but 2RX Read IOPS.  This is dramatic.  Suddenly when viewing performance as a factor of stable capacity we find that RAID 10 has double read performance over RAID 0!

RAID 5 gets slightly trickier.  Write IOPS would be expressed as ((R + 1) * X)/4.  The Read IOPS are expressed as ((R +1) * X).

RAID 6, as we expect, follows the pattern that RAID 5 projects.  Write IOPS for RAID 6 are ((R + 2) * X)/6.  And the Read IOPS are expressed as ((R + 2) * X).

RAID 7 falls right in line.  RAID 7 Write IOPS would be ((R + 3) * X)/8.  And the Read IOPS are ((R + 3) * X).

This vantage point changes the way that we think about performance and, when looking purely at read performance, RAID 0 becomes the slowest RAID level rather than the fastest and RAID 10 becomes the fastest for both read and write no matter what the values are for R and X!

If we take a real world example of 10 2TB drives to achieve 20TB of usable capacity with each drive having 100 IOPS of performance and assume a 50/50 blend, the resultant IOPS would be:  RAID 0 with 1,000 Blended IOPS, RAID 10 with 1,500 Blended IOPS (2,000 RIOPS / 1,000 WIOPS), RAID 5 with 687.5 Blended IOPS (1,100 RIOPS / 275 WIOPS), RAID 6 with 700 Blended IOPS (1,200 RIOPS / 200 WIOPS) and finally RAID 7 with 731.25 Blended IOPS (1,300 RIOPS / 162.5 WIOPS.)  RAID 10 is a dramatic winner here.

Latency and System Impact with Software RAID

As I have stated earlier, RAID 0 and RAID 10 have, effectively, no system overhead to consider.  The mirroring operation requires essentially no computational effort and is, for all intents and purposes, immeasurably small.  Parity RAID does have computational overhead and this results in latency at the storage layer and system resources being consumed.  Of course, if we are using hardware RAID those resources are dedicated to the RAID array and have no function but to be consumed in this role.  If we are using software RAID, however, these are general purpose system resources (primarily CPU) that are consumed for the purposes of the RAID array processing.

The impact to even a very small system with a large amount of RAID is still very small but it can be measured and should be considered, if only lightly.  Latency and system impact are directly related to one another.

There is no simple way to state latency and system impact for different RAID levels except in this way: RAID 0 and RAID 10 have effectively no latency or impact, RAID 5 has some latency and impact, RAID 6 has roughly twice as much computational latency and impact as RAID 5 and RAID 7 has roughly triple the computational latency and impact as RAID 5.

In many cases this latency and system impact will be so small that they cannot be measured with standard system tools and as modern processors become increasingly powerful the latency and system impact will continue to diminish.  Impact has been considered negligible for RAID 5 and RAID 6 systems on even low end, commodity hardware since approximately 2001.  But it is possible on heavily loaded systems with a large amount of parity RAID activity that there could be contention between the RAID subsystem and other processes requiring system resources.

Reference: The IT Hollow – Understanding the RAID Penalty

Article originally posted to the StorageCraft Blog – RAID Performance.