Louwrentius

Introduction

The goal of this blogpost is to help you better understand storage performance. I want to discuss some fundamentals that are true regardless of your particular needs.

This will help you better reason about storage and may provide a scaffolding for further learning.

If you run your applications / workloads entirely in the cloud, this information may feel antiquated or irrelevant.

However, since the cloud is just somebody else's compute and storage, knowledge about storage may still be relevant. Cloud providers expose storage performance metrics for you to monitor and this may help to make sense of them.

Concepts

I/O

An I/O is a single read/write request. That I/O is issued to a storage medium (like a hard drive or solid state drive).

It can be a request to read a particular file from disk. Or it can be a request to write some data to an existing file. Reading or writing a file can result in multiple I/O requests.

I/O Request Size

The I/O request has a size. The request can be small (like 1 Kilobyte) or large (several megabytes). Different application workloads will issue I/O operations with different request sizes. The I/O request size can impact latency and IOPS figures (two metrics we will discuss shortly).

IOPS

IOPS stands for I/O Operations Per Second. It is a performance metric that is used (and abused) a lot in the world of storage. It tells us how many I/O requests per second can be handled by the storage (for a particular workload).

Warning: this metric is meaningless without a latency figure. We will discuss latency shortly.

Bandwidth or throughput

If you multiply the IOPS figure with the (average) I/O request size, you get the bandwidth or throughput. We state storage bandwidth mostly in Megabytes and Gigabytes per second.

To give you an example: if we issue a workload of 1000 IOPS with a request size of 4 Kilobytes, we will get a throughput of 1000 x 4 KB = 4000 KB. This is about ~4 Megabytes per second.

Latency

Latency is the time it takes for the I/O request to be completed. We start our measurement from the moment the request is issued to the storage layer and stop measuring when either we get the requested data, or get confirmation that the data is stored on disk.

Latency is the single most important metric to focus on when it comes to storage performance, under most circumstances.

For hard drives, an average latency somewhere between 10 to 20 ms is considered acceptable (20 ms is the upper limit).

For solid state drives, depending on the workload it should never reach higher than 1-3 ms. In most cases, workloads will experience less than 1ms latency numbers.

IOPS and Latency

This is a very important concept to understand. The IOPS metric is meaningless without a statement about latency. You must understand how long each I/O operation will take because latency dictates the responsiveness of individual I/O operations.

If a storage solution can reach 10,000 IOPS but only at an average latency of 50 ms that could result in very bad application performance. If we want to hit an upper latency target of 10 ms the storage solution may only be capable of 2,000 IOPS.

For more details on this topic I would recommend this blog and this blog.

Access Patterns

Sequential access

An example of a sequential data transfer is copying a large file from one hard drive to another. A large number of sequential (often adjacent) datablocks is read from the source drive and written to another drive. Backup jobs also cause sequential access patterns.

In practice this access pattern shows the highest possible throughput.

Hard drives have it easy as they don't have to spend much time moving their read/write heads and can spend most time reading / writing the actual data.

Random access

I/O requests are issued in a seemingly random pattern to the storage media. The data could be stored all over various regions on the storage media. An example of such an access pattern is a heavy utilised database server or a virtualisation host running a lot of virtual machines (all operating simultaneously).

Hard drives will have to spend a lot of time moving their read/write heads and can only spend little time transferring data. Both throughput and IOPS will plummet (as compared to a sequential access pattern).

In practice, most common workloads, such as running databases or virtual machines, cause random access patterns on the storage system.

Queue depth

The queue depth is a number between 1 and ~128 that shows how many I/O requests are queued (in-flight) on average. Having a queue is beneficial as the requests in the queue can be submitted to the storage subsystem in an optimised manner and often in parallel. A queue improves performance at the cost of latency.

If you have some kind of storage performance monitoring solution in place, a high queue depth could be an indication that the storage subsystem cannot handle the workload. You may also observe higher than normal latency figures. As long as latency figures are still within tolerable limits, there may be no problem.

Storage Media Performance characteristics

Hard drives

Hard drives (HDDs) are mechanical devices that resemble a record player.

They have an arm with a read/write head and the data is stored on (multiple) platters.

Hard drives have to physically move read/write heads to fulfil read/write requests. This mechanical nature makes them relatively slow as compared to solid state drives (which we will cover shortly).

Especially random access workloads cause hard drives to spend a lot of time on moving the read/write head to the right position at the right time, so less time is available for actual data transfers.

The most important thing to know about hard drives is that from a performance perspective (focussing on latency) higher spindle speeds reduce the average latency.

Because the latency of individual I/O requests is lower the drives with a higher RPM, you can issue more of such requests in the same amount of time. That's why the IOPS figure also increases.

Latency and IOPS of an older Western Digital Velociraptor 10,000 RPM drive:

Notice the latency and IOPS in the Queue Depth = 1 column.

Source used to validate my own research.

Regarding sequential throughput we can state that fairly old hard drives can sustain throughputs of 100-150 megabytes per second. More modern hard drives with higher capacities can often sustain between 200 - 270 megabytes per second.

An important note: sequential transfer speeds are not constant and depend on the physical location of the data on the hard drive platters. As a drive fills up, throughput diminishes. Throughput can drop more than fifty percent! ¹.

So if you want to calculate how long it will take to transfer a particular (large) dataset, you need to take this into account.

Solid State Drives

Solid state drives (SSDs) have no moving parts, they are based on flash memory (chips). SSDs can handle I/O much faster and thus show significantly lower latency.

Whereas we measure the average I/O latency of HDDs in milliseconds (a thousand of a second) we measure the latency of SSD I/O operations in microseconds (a millionth of a second).

Because of this reduced latency per I/O request, SSDs outperform HDDs in every conceivable way. Even a cheap consumer SSD can at least sustain about 5000+ IOPS with only a 0.15 millisecond (150 microseconds) latency. That latency is about 40x better than the best latency of an enterprise 15K RPM hard drive.

Solid state drives can often handle I/O requests in parallel. This means that larger queue depths with more I/O requests in flight can show significantly higher IOPS with a limited (but not insignificant) increase in latency.

The random I/O performance of an older SATA consumer SSD

More modern enterprise SSDs show better latency and IOPS. The SATA interface seems the main bottleneck.

The random I/O performance of an enterprise SATA SSD

SSDs perform better than HDDs across all relevant metrics except price in relation to capacity.

Important note: SSDs are not well-suited for archival storage of data. Data is stored as charges in the chips and those charges can diminish over time. It's expected that even hard drives are better suited for offline archival purposes although the most suitable storage method would probably be tape.

SSD actual performance vs advertised performance

Many SSDs are advertised with performance figures of 80,000 - 100,000 IOPS at some decent latency. Depending on the workload, you may only observe a fraction of that performance.

Most of those high 80K-100K IOPS figures are obtained by benchmarking with very high queue depths (16-32). The SSD benefits from such queue depths because it can handle a lot of those I/O requests in parallel.

Please beware: if your workload doesn't fit in that pattern, you may see lower performance numbers.

If we take a look at the chart above of the Intel SSD we may notice how the IOPS figures only start to come close to the advertised 80K+ IOPS as the queue depth increases. It's therefore important to understand the characteristics of your own workload.

RAID

If we group several hard drives together we can create a RAID array. A RAID array is a virtual storage device that exceeds the capacity and performance of a single hard drive. This allows storage to scale within the limits of a single computer.

RAID is also used (or some say primarily used) to assure availability by assuring redundancy (drive failure won't cause data loss). But for this article we focus it's performance characteristics.

SSDs can achieve impressive sequential throughput speeds, of multiple gigabytes per second. Individual hard drives can never come close to those speeds, but if you put a lot of them together in a RAID array, you can come very close. For instance, my own NAS an achieve such speeds using 24 drives.

RAID also improves the performance of random access patterns. The hard drives in a RAID array work in tandem to service those I/O requests so a RAID array shows significantly higher IOPS than a single drive. More drives means more IOPS.

RAID 5 with 8 x 7200 RPM drives

The picture below shows the read IOPS performance of an 8-drive RAID 5 array of 1 TB, 7200 RPM drives. We run a benchmark of random 4K read requests.

Notice how the IOPS increases as the queue depth increases.

However, nothing is free in this world. A higher queue depth - which acts as a buffer - does increase latency.

Notice how quickly the latency exceeds 20ms and quickly becomes almost unusable.

RAID 5 with 8 x 10,000 RPM drives

Below is the result of a similar test with 10,000 RPM hard drives. Notice how much better the IOPS and latency figures are.

The latency looks much better:

It makes sense to put SSDs in RAID. Although they are more reliable than hard drives, they can fail. If you care about availability, RAID is inevitable. Furthermore, you can observe the same benefits as with hard drives: you pool resources together, achieving higher IOPS figures and more capacity than possible with a single SSD.

Capacity vs. Performance

The following is mostly focussed on hard drives although it could be true for solid state drives as well.

We put hard drives in RAID arrays to get more IOPS than a single drive can provide. At some point - as the workload increases - we may hit the maximum number of IOPS the RAID array can sustain with an acceptable latency.

This IOPS/Latency threshold could be reached even if we have only 50% of the storage capacity of our RAID array in use. If we use the RAID array to host virtual machines for instance, we cannot add more virtual machines because this would cause the latency to rise to unacceptable levels.

It may feel like a lot of good storage space is going to waste, and in some sense this may be true. For this reason, it could be a wise strategy to buy smaller 10,000 RPM or 15,000 RPM drives purely for the IOPS they can provide and forgo on capacity.

So it might be the case that you may have to order and add let's say 10 more hard drives to meet the IOPS/Latency demands while there's still plenty of space left.

This kind of situation is less likely as SSDs have taken over the role of the performance storage layer and (larger capacity) hard drives are pushed in the role of 'online' archival storage.

Closing words

I hope this article has given you a better understanding of storage performance. Although it is just an introduction, it may help you to better understand the challenges of storage performance.

Introduction

Update: I've noticed some strange behavior of SSDs when benchmarking them with FIO. After further investigation and additional testing, I've found the reason for the strange patterns in the graphs.

The 'strange' test results are due to the fact that they were obtained by connecting the SSDs to a P420I controller. As the HBA mode of this controller performs worse than the RAID mode, I used the RAID mode of this controller. Indvidual drives were put in a RAID0 volume. But it turns out that this creates a strange interaction between RAID controller and SSD.

Additional testing with an SATA 300 AHCI controller shows 'normal' patterns that look similar to the results of the INTEL SSD as compared to the other ones (Samsung and Kingston).

It seems I've made a mistake by using the P420i controller for testing. I have includes both 'bad' and 'good' results.

Regular SATA solid state drives may seem interchangeable at this point. They all show amazing IOPS and latency performance.

I have performed benchmarks on different SSDs from different vendors and it seems that they actually show very different behaviour. This behavior has come to light because I benchmarked the entire device capacity.

The benchmark - performed with FIO - puts a fifty percent read/write random 4K workload on the device. The benchmark stops when all sectors of the device have been read or written to. Furthermore, all tests are performed with a queue depth of 1.

I've made this post because I found the results interesting. At least the images show a very peculiar pattern for some SSDs. I can't explain them really, maybe you can.

This is the test I ran against the SSDs.

fio --filename=/dev/sdX --direct=1 --rw=randrw --refill_buffers
--norandommap --ioengine=libaio --bs=4k --rwmixread=50 --iodepth=1

Disclaimer

I've performed these benchmark to the best of my knowledge. The raw benchmark data is available here

It's always possible that I made a mistake, so it may be wise to run your own tests to see if you can replicate these results.

Caveat: I really don't know if these benchmark results impact real-life performance. Maybe these benchmark results show a kind of behaviour of SSDs that doesn't really matter in the end.

Benchmark Results

Intel D3-S4610

This SSD is meant for for datacenter usage. This is the test result on the P420i controller.

IOPS and Latency is consistent during the whole benchmark. It's behaviour seems predictable.

This is the test result on the AHCI controller:

Samsung 860 Pro

This SSD is meant for desktop usage. Its behavior seems quite different from the Intel SSD. I have separated the IOPS data from the Latency data to make the graphs more eligible.

This is the test result on the P420i controller.

IOPS

Latency

The best-case latency is almost four times better than the worst-case latency. Latency is thus less predictable. This impact also seems to be reflected in the IOPs numbers.

This is the test result on the AHCI controller:

Samsung PM883

This SSD is meant for datacenter usage. This is the test result on the P420i controller.

IOPS

Latency

This SSD seems to behave in a similar way as the 860 PRO.

This is the test result on the AHCI controller:

Kingston DC500M

This SSD is meant for datacenter usage. This is the test result on the P420i controller.

IOPS

latency

The behavior of this SSD seems similar to the behaviour of the Samsung SSDs but the pattern is distinct: it seems shifted as compared to the Samsung SSDs.

This is the test result on the AHCI controller:

Evaluation

** Updated evaluation**

We can conclude that the P420i RAID controller causes strange behavior not observed when we test the SSDs on a regular AHCI controller. Although this was an older SATA 300 controller, I'm making the assumption that this controller still has enough bandwidth to support a random 4K test as most tests never went beyond 50+ MB/s of throughput.

At this point, I can only say that I observe quite different behavior between the Intel SSD and the other SSDs from Samsung and Kingston. The problem is that I can't tell if this affects real-life day-to-day application performance.

It seems that although results for the Samsung and Kingston SSDs fluctuate quite a bit, it's quite possible that the fluctuations occur during a very short timespan and effectively cancel each other out.

If you have comments, ideas or suggestions, leave a comment below.

How are these images generated?

All images have been generated with fio-plot.

The github repository also contains a folder with a lot of example images.

Introduction

To learn more about Ceph, I've build myself a Ceph Cluster based on actual hardware. In this blogpost I'll discus the cluster in more detail and I've also included (fio) benchmark results.

This is my test Ceph cluster:

The cluster consists of the following components:

 3 x Raspberry Pi 3 Model B+ as Ceph monitors
 4 x HP MicroServer as OSD nodes (3 x Gen8 + 1 x Gen10)
 4 x 4 x 1 TB drives for storage (16 TB raw)
 3 x 1 x 250 GB SSD (750 GB raw)
 2 x 5-port Netgear switches for Ceph backend network (bonding)

Monitors: Raspberry Pi 3 Model B+

I've done some work getting Ceph compiled on a Raspberry Pi 3 Model B+ running Raspbian. I'm using three Raspberry Pi's as Ceph monitor nodes. The Pi boards don't break a sweat with this small cluster setup.

Note: Raspberry Pi's are not an ideal choice as a monitor node because Ceph Monitors write data (probably the cluster state) to disk every few seconds. This will wear out the SD card eventually.

Storage nodes: HP MicroServer

The storage nodes are based on four HP MicroServers. I really like these small boxes, they are sturdy, contain server-grade components, including ECC-memory and have room for four internal 3.5" hard drives. You can also install 2.5" hard drives or SSDs.

For more info on the Gen8 and the Gen10 click on their links.

Unfortunately the Gen8 servers are no longer made. The replacement, the Gen10 model, lacks IPMI/iLO and is also much more expensive (in Europe at least).

CPU and RAM

All HP Microservers have a dual-core CPU. The Gen8 servers have 10GB RAM and the Gen10 server has 12GB RAM. I've just added an 8GB ECC memory module to each server, the Gen10 comes with 4GB and the Gen8 came with only 2GB, which explains the difference.

Boot drive

The systems all have an (old) internal 2.5" laptop HDD connected to the internal USB 2.0 header using an USB enclosure.

Ceph OSD HDD

All servers are fitted with four (old) 1TB 7200 RPM 3.5" hard drives, so the entire cluster contains 16 x 1TB drives.

Ceph OSD SSD

There is a fifth SATA connector on the motherboard, meant for an optional optical drive, which I have no use for and wich is not included with the servers.

I use this SATA connector in the Gen8 MicroServers to attach a Crucial 250GB SSD, which is then tucked away at the top, where the optical drive would sit. So the Gen8 servers have an SSD installed which the Gen10 is lacking.

The entire cluster thus has 3 x 250GB SSDs installed.

Networking

All servers have two 1Gbit network cards on-board and a third one installed in one of the half-height PCIe slots¹.

The half-height PCIe NICs connect the Microservers to the public network. The internal gigabit NICs are configured in a bond (round-robin) and connected to two 5-port Netgear gigabit switches. This is the cluster network or the backend network Ceph uses for replicating data between the storage nodes.

You may notice that the first onboard NIC of each server is connected to the top switch and the second one is connected to the bottom switch. This is necessary because linux round-robin bonding requires either separate VLANs for each NIC or in this case separate switches.

Benchmarks

Benchmark conditions

• The tests ran on a physical Ceph client based on an older dual-core CPU and 8GB of RAM. This machine was connected to the cluster with a single gigabit network card.
• I've mapped RBD block devices from the HDD pool and the SSD pool on this machine for benchmarking.
• All tests have been performed on the raw /dev/rbd0 device, not on any file or filesystem.
• The pools use replication with a minimal copy count of 1 and a maximum of 3.
• All benchmarks have been performed with FIO.
• All benchmarks used random 4K reads/writes

    NAME     ID     USED        %USED     MAX AVAIL     OBJECTS
    hdd      36     1.47TiB     22.64       5.03TiB      396434
    ssd      38      200GiB     90.92       20.0GiB       51204

Benchmark SSD

Click on the images below to see a larger version.

Benchmark HDD

Benchmark evaluation

The random read performance of the hard drives seems unrealistic at higher queue depths and number of simultaneous jobs. This performance cannot be sustained purely on the basis that 16 hard drives with maybe 70 random IOPs each can only sustain 1120 random IOPs.

I cannot explain why I get these numbers. If anybody has a suggestion, feel free to comment/respond. Maybe the total of 42GB of memory across the cluster may act as some kind of cache.

Another interesting observation is that a low number of threads and a small IO queue depth results in fairly poor performance, both for SSD and HDD media.

Especially the performance of the SSD pool is poor with a low IO queue depth. A probable cause is that these SSDs are consumer-grade and don't perform well with low queue depth workloads.

I find it interesting that even over a single 1Gbit link, the SSD-backed pool is able to sustain 20K+ IOPs at higher queue depths and larger number of threads.

The small number of storage nodes and the low number of OSDs per node doesn't make this setup ideal but it does seem to perform fairly decent, considering the hardware involved.