Following on from the previous [1] [2] experiments with Postgres & pgbench. A quick look at how the workload is seen from the Nutanix CVM.
The Linux VM running postgres has two virtual disks:
One is taking transaction log writes.
The other is doing reads and writes from the main datafiles.
Since the database size is small (50% the size of the Linux RAM) – the data is mostly cached inside the guest, and so most reads do not hit storage. As a result we only see writes going to the DB files.
Additionally, we see that database datafile writes the arrive in a bursty fashion, and that these write bursts are more intense (~10x) than the log file writes.
Charts from Prometheus/Grafana showing IO rates seen from the perspective of the Linux guest VM
Despite the database flushes ocurring in bursts with a decent amount of concurrency the Nutanix CVM provides an average of 1.5ms write response time.
From the Nutanix CVM port 2009 handler, we can access the individual vdisk statistics. In this particular case vDisk 45269 is the data file disk, and 40043 is the database transaction log disk.
Datafile writes completed in 1.5millisecond average – despite deep queues during burst
The vdisk categorizer correctly identifies the database datafile write pattern as highly random.
Writes to the datbase datafiles are almost entirely random
As a result, the writes are passed into the replicated oplog
The burst of writes hits the oplog as expected
Meanwhile the log writes are categorized as mostly sequential, which is expected for a database log file workload.
Meanwhile, log file writes are mostly categorized as sequential.
Even though the log writes are sequential, they are low-concurrency and small size (looks like mostly 16K-32K). This write pattern is also a good candidate for oplog.
Once connected to the appliance we can use postgres and pgbench to generate simplistic database workload.
[1] Do this on a Linux box somewhere. For some reason the conversion failed using my qemu utilities installed via brew. Importing OVAs direct into AHV should be available in the future.
Nutanix AOS 5.10 ships with a feature called Autonomous Extent Store (AES). AES effectively provides Metadata Locality to complement the existing data locality that has always existed. For large datasets (e.g. a 10TB database with 20% hot data) we observe a 2X improvement in throughput for random access across the 2TB hot dataset.
In our experiment we deliberately size the active working-set to NOT fit into the metadata cache. We uniformly access 2TB with a 100% random access pattern and record the time to access all 2TB. On the same hardware with AES enabled – the time is cut in half. As can be seen in the chart – the throughput is double, as expected.
It is the localization of metadata from AES that contributes to the 2X improvement. AES keeps most of the metadata local to the node – so there is no need to fetch data across-the-wire. Additionally AES reduces the need to cache metadata in DRAM since local access is so fast. For very large datasets, retrieving metadata can contribute a large proportion of the access time. This is true for all storage, so speeding up metadata resolution can make a dramatic improvement to overall throughput as we demonstrate.
During .Next 2018 in London, Nutanix announced performance improvements in the core-datapath said to give up to 2X performance improvements. Here’s a real-world example of that improvement in practice.
I am using X-Ray to simulate a 1TB data restore into an existing database. Specifically the IO sizes are large, an even split of 64K,128K,256K, 1MB and the pattern is 100% random across the entire 1TB dataset.
bssplit=64k/20:128k/20:256k/20:512k/20:1m/20
Normally storage benchmarks using large IO sizes are performed serially, because it’s easier on the storage back-end. That may be realistic for an initial load, but in this case we want to simulate a restore where the pattern is 100% random.
In this case the time to ingest 1TB drops by half when using Nutanix AOS 5.10 with Autonomous Extent Store (AES) enabled Vs the previous traditional extent store.
This improvement is possible because with AES, inserting directly into the extent store is much faster.
For throughput sensitive, random workloads, AES can detect that it will be faster to skip the oplog. Skipping oplog allows AES to eliminate a network round trip to a remote oplog – and instead only make an RF2 copy for the Extent Store. By contrast, when sustained, large random IO is funneled into oplog, the 10Gbit network can become the bottleneck. Even with faster networks, AES will still be a benefit because the CPU and SSD resource usage is also lower. Unfortunately I only have 10Gbit networking in my lab!
The X-Ray files needed to run this test are on github
In a previous post I showed a chart which plots concurrency [X-axis] against throughput (IOPS) on the Y-Axis. Here is that plot again:
Experienced performance chart ogglers will notice the familiar pattern of Littles Law, whereby throughput (X) rises quickly as concurrency (N) is increased. As we follow the chart to the right, the slope flattens out and we achieve a lower increase in throughput, even as we increase concurrency by the same amount at each stage. The flattening of the curve is best understood as Amdahls Law.
Anyone who follows Dr. Neil Gunther and his Universal Scalability Law, will also recognize this curve.
The USL states that taking the values of concurrency and throughput as inputs, we can in fact calculate the scalability of the system. Specifically we are able to calculate the key factors of contention and crosstalk – which limit absolute linear scalability and eventually result in less throughput as additional load is submitted even as the capacity of the system is saturated.
Using his Excel spreadsheet, I was able to input the numbers from my test and derive values that determine scalability.
Taking the largest number (0.074%) the “contention value” (i.e the impact we expect due to Amdahls law) as the limit to linear scaling – we can say that for this particular cluster, running this particular (simplistic/synthetic) workload the Nutanix cluster scales 99.926% linear. Although I did not crank up the concurrency beyond 576, the model shows us that this cluster will start to degrade performance if we try to push concurrency beyond 600 or so. Again, the USL model is for this particular workload – on this particular cluster. Doubling the concurrency of the offered load to 1200 will only net us 500,000 IOPS according to the model.
The high linearity (99.926%) is expected. The workload is 100% read, and with the data-locality feature of Nutanix filesystem – we expect close to 100% scalability.
We will return to these measures of scalability in the future to look at more realistic workloads.
The fio Pareto parameter allows us to create a workload, which references a very large dataset, but specify a hotspot for the access pattern. Here’s an example using the same setup as the ILM experiment, but using a Pareto value of 0:8. My fio file looks like this..
[global]
ioengine=libaio
direct=1
time_based
norandommap
random_distribution=pareto:0.8 The experiment shows that with the access pattern as a Pareto ratio 0:8, meaning 20% of the overall dataset is “hot” the ILM process happens much faster as the hotspot is smaller, and is identified faster than a 100% uniform random access pattern. We would expect a similar shape for any sort of caching mechanism.
At some point potential Hyper-converged infrastructure (HCI) users want to know – “How fast does this thing go?”. The real question is “how do we measure that?”.
The simplest test is to run a single VM, with a single disk and issue a single IO at a time. We see often see this sort of test in bake-offs, and such a test does answer an important question – “what’s the lowest possible response time I can expect from the storage”.
However, this test only gives a single data point. Since nobody purchases a HCI cluster to run a single VM, we also need to know what happens when multiple VMs are run at the same time. This is a much more difficult test to conduct, and many end-users lack access and experience with tools that can give the full picture.
In the example below, the single VM, single vdisk, single IO result is at the very far left of the chart. Since it’s impossible to read I will tell you that the result is about 2,500 IOPS at ~400 microseconds. (in fact we know that if the IOPS are 2,500 the response time MUST be 400 microseconds 1/2,500 == .0004 seconds)
However with a single VM, the cluster is mostly idle, and has capacity to do much more work. In this X-Ray test I add another worker VM doing the exact same workload pattern to every node in the cluster every 5 minutes.
By the time we reach the end of the test, the total IOPS have increased to around 600,000 and the response time only increased by an additional 400 microseconds.
In other words the cluster was able to achieve 240X the amount of work measured by the single VM on a single node with only a 2X increase in response time, which is still less than 1ms.
The overall result is counter-intuitive, because the rate of change in IOPS (240X) is way out of line with the increase in response time (2X). The single VM test is using only a fraction of the cluster capacity.
When comparing HCI clusters to traditional storage arrays – you should expect the traditional array to outperform the cluster at the far left of the chart, but as work scales up the latent capacity of the HCI cluster is able to provide huge amounts of IO with very low response times.
You can run this test yourself by adding this custom workload to X-Ray
Specifically a customer wanted to see how performance changes (and how quickly) as data moves from HDD to SSD automatically as data is accessed. The access pattern is 100% random across the entire disk.
In a hybrid Flash/HDD system – “cold” data (i.e. data that has not been accessed for a long time) is moved from SSD to HDD when the SSD capacity is exhausted. At some point in the future – that same data may become “hot” again, and so we want to make sure that the “newly hot” data is quickly moved back to the SSD tier. The duration of the above chart is around 5 minutes – and we see that by, around the 3 minute mark the entire dataset is resident on the SSD tier.
This X-ray test uses a couple of neat tricks to demonstrate ILM behavior.
Edit container preferences to send sequential data immediately to HDD
Overwriting data with NUL/Zero bytes frees the underlying data on Nutanix filesystem
To demonstrate ILM from HDD to SSD (ad ultimately into the DRAM cache on the CVM) we first have to ensure that we have data on the HDD in the first place. By default Nutanix OS will always try to write new data to SSD. To circumvent that behavior we can edit the container preferences. We use the fact that the “prefill” will be a sequential workload, while the measured workload will be a random workload. To make the change, use “ncli” to change the ” Sequential I/O Pri Order” to be HDD only.
In my case I happened to call my container “xray” since I didn’t want to change the default container. Now, when X-Ray executes the prefill stage, the data will land on HDD. As a second requirement, we want to see what happens when IO with different size blocks are issued so that we can get a chart similar to this: To achieve the desired behavior, we need to make sure that, at the beginning of each test, the data, again resides on HDD. The problem is that the data is up-migrated during the test. To do this we do an initial overwrite of the entire disk with “NULL” bytes using a parameter in fio “zero_buffers”. This causes the data to be freed on the Nutanix filesystem. Then we issue a normal profile with random data. Once the data is freed, then we know that the new initial writes will go to HDD – because we edited the container to do so. The overall test pattern looks like this
Create and clone VMs
Prefill with random data (Data will reside on HDD due to container edit)
Read disk with 16KB block size
Zero out the disks – to remove/free the up-migrated data
What happens when power is lost to all nodes of a HCI Cluster?
Ever wondered what happens when all power is simultaneously lost on a HCI cluster? One of the core principles of cloud design is that components are expected to fail, but the cluster as a whole should stay “up”. We wanted to see what happens when all components fail at once, so we designed an X-Ray test to do exactly that.
We start an OLTP workload on every node in the cluster, then X-Ray connects to the IPMI port on each node, and powers off all the hosts while the cluster is under load. In particular, the cluster is under read/write load (we need write workload, because we want to force the cluster to recover in-flight writes).
After power-off, we wait 10 seconds for everything to spin down, then immediately re-apply the power by connecting to the IPMI ports.
The nodes power up, and immediately start their POST (Power On Self Test) and boot the hypervisor. The CVM will auto-start, discover the available nodes and form the cluster.
X-Ray polls the cluster manager (either Prism or vCenter) to determine that the cluster is “up” and then restarts the OLTP workload.
Our testing showed that our Nutanix cluster completed POST, and was ready to restart work in around 10 minutes. Moreover, the time to achieve the recovery had very little variability. The chart below shows three separate runs on the same cluster.
This is the YAML file which defines the workload. The full specification is on github. The key part of the YAML is the nodes.PowerOff which connects to the IMPI ports of each node and vm_group.WaitForPowerOn which connects to either Nutanix Prism or vmware vCenter and determines that the cluster is formed, and ready to accept new work.
Creating a HCI benchmark to simulate multi-tennent workloads
HCI deployments are typically multi-tennant and often different nodes will support different types of workloads. It is very common to have large resource-hungry databases separated across nodes using anti-affinity rules. As with traditional storage, applications are writing to a shared storage environment which is necessary to support VM movement. It is the shared storage that often causes performance issues for data bases which are otherwise separated across nodes. We call this the noisy neighbor problem. A particular problem occurs when a reporting / analytical workload shares storage with a transactional workload.
In such a case we have a Bandwidth heavy workload profile (reporting) sharing with a Latency Sensitive workload (transactional)
In the past it has been difficult to measure the noisy neighbor impact without going to the trouble of configuring the entire DB stack, and finding some way to drive it. However in X-Ray we can do exactly this sort of workload. We supply a pre-configured scenario which we call the DB Colocation test.
The DB Colocation test utilizes two properties of X-Ray not found in other benchmarking tools
Time based benchmark actions
Distinct per-VM workload patterns
Ability to provision particular workloads, to particular hosts
In our example scenarioX-Ray begins by starting a workload modeled after a transactional DB (we call this the OLTP workload) on one of the nodes. This workload runs for 60 minutes. Then after 30 minutes X-Ray starts workloads modeled after reporting/analytical workloads on two other nodes (we call this the DSS workload).
After 30 minutes we have three independent workloads running on three independent nodes, but sharing the same storage. The key thing to observe is the impact on the latency sensitive (OLTP) workload. In this experiment it is the DSS workloads which are the noisy neighbor, since they will tend to utilize a lot of the storage bandwidth. An ideal result is one where there is very little interference with the running OLTP workload, even though we expect latency to increase. We can compare the impact on the OLTP workload by comparing the IOPS/response time during the first 30 minutes (no interference) with the remaining 60 minutes (after the DSS workloads are started). We should expect to see some increase in response time from the OLTP application because the other nodes in the cluster have gone from idle to under-load. The key thing to observe is whether the OLTP IOP target rate (4,000 IOPS) is achieved when the reporting workload is applied.
X-Ray Scenario configuration
We specify the timing rules and workloads in the test.yml file. You can modify this to contain whichever values suit your model. I covered editing an existing workload in Part 1.
The overall scenario begins with the OLTP workload, which will run for 3600 seconds (1 hour). The stagger_secs value is used if there are multiple OLTP sub-workloads. In the simple case we do use a single OLTP workload.
The scenario pauses for 1800 seconds using the test.wait specification then immediately starts the DSS workload
Finally the scenario uses the workload.Wait specification to wait for the OLTP workload to finish (approx 1 hour) before the test is deemed completed.
X-Ray Workload specification
The DB Co-Location test uses two workload profiles that aim to simulate transactional (OLTP) and reporting/analytical (DSS) workloads. The specifications for those workloads are contained in the two .fio files (oltp.fio and dss.fio)
OLTP
The OLTP workload (oltp.fio) that we ship as has the following characteristcs based on typical configurations that we see in the field (of course you can change these to whatever you like).
Target IOP rate of 4,000 IOPS
4 “Data” Disks
50/50 read/write ratio.
90% 8KB, 10% 32KB bloc-ksize
8 outstanding IO per disk
2 “Log” Disks
100% write
90% sequential
32k block-size
1 outstanding IO per disk
The idea here is to simulate the two main storage workloads of a DB. The “data” portion and the “log” portion. Log writes are just used to commit transactions and so are 100% write. The only time the logs are read is during DB recovery, which is not part of this scenario. The “Data” disks are doing both reads (from DB cache misses) and writes committed transactions. A 50/50 read/write mix might be considered too write intensive – but we wanted to stress the storage in this scenario.
DSS
The DSS workload is configured to have the following characteristics
Target IOP rate of 1400 IOPS
4 “Data” Disks
100% Read workload with 1MB blocksize
10 Outstanding IOs
2 “Log” Disk
100% Write workload
90% sequential
32K block-size
1 outstanding IO per disk
The idea here is to simulate a large database doing a lot of reads across a large workingset size. The IO to the data disks is entirely read, and uses large blocks to simulate a database scanning a lot of records. The “Log” disks have a very light workload, purely to simulate an active database which is probably updating a few tables used for housekeeping.
It’s good to detect corrupted data. It’s even better to transparently repair that data and return the correct data to the user. Here we will demonstrate how Nutanix filesystem detects and corrects corruption. Not all systems are made equally in this regard. The topic of corruption detection and remedy was the focus of this excellent Usenix paper Redundancy Does Not Imply Fault Tolerance: Analysis of Distributed Storage Reactions to Single Errors and Corruptions. The authors find that many systems that should in theory be able to recover corrupted data do not in fact do so.
Within the guest Virtual Machine
Start with a Linux VM and write a specific pattern (0xdeadbeef) to /dev/sdg using fio.
Check that the expected data is written to the virtual disk and generate a SHA1 checksum of the entire disk.
The “od” command shows us that the entire 1GB disk contains the pattern 0xdeadbeef
The “sha1sum” command creates a checksum (digest) based on the content of the entire disk.
Within the Nutanix CVM
Connect to the Nutanix CVM
Locate one of the 4MB egroups that back this virtual disk on the node.
The virtual disk which belongs to the guest vm (/dev/sdg) is represented in the Nutanix cluster as a series of “Egroups” within the Nutanix filesystem.
Using some knowledge of the internals I can locate the Egroups which make up the vDisk seen by the guest.
Double check that this is indeed an Egroup belonging to my vDisk by checking that it contains the expected pattern (0xdeadbeef)
Now simulate a hardware failure and overwrite the egroup with null data
I do this by reaching underneath the cluster filesystem and deliberately creating corruption, simulating a mis-directed write somewhere in the system.
If the system does not correct this situation, the user VM will not read 0xdeadbeef as it expects – remember the corruption happened outside of the user VM itself.
As an experiment, I wanted to (a) Create a HDD only container, and (b) measure the bandwidth I could achieve when backing up the SQL DB. This was performed on a standard hybrid platform with only 4 HDD’s in the node.
First create a container, but add the special options “sequential-io-priority-order=DAS-SATA random-io-priority-order=DAS-SATA” which means that all IO will be directed to the HDD only. This also means that data on this container will never be migrated up. This is just fine for a backup that will hopefully never be read, and if it is – only once, sequentially.
Next create a vDisk in that container – this disk will contain the SQL Server backup data
Add vdisk to the cold-only container.
Format and initialize the drive.
Format the drive to hold SQL backup.
Add backup targets to the drive. Adding multiple targets increases throughput because SQL Server will generate 1-2 outstanding IO’s per target. I created 16 total targets (these are just files).
The first backup is a little slow (~64MB/s), because we’re creating the files. A second (and subsequent) backups go faster, around 120 MB/s writing directly to the HDD spindles on a single node with 4 HDDs.
This backup stream drives around 25MB/s per HDD spindle on the Nutanix node. On a larger platform with more spindles – we could easily drive 500MB/s, and still skip SSD by writing directly to HDD.
Backup just started. About 115MB/s read, 115MB/s write on same node.
The question of why Nutanix uses SATA drive comes up sometimes, especially from customers who have experienced very poor performance using SATA on traditional arrays.
I can understand this anxiety. In my time at NetApp we exclusively used SAS or FC-AL drives in performance test work. At the time there was a huge difference in performance between SCSI and SATA. Even a few short years ago, FC typically spun at 15K RPM whereas SATA was stuck at about a 5K RPM, so experiencing 3X the rotational delay.
These days SAS and SATA are both available in 7200 RPM configurations, and these are the type we use in standard Nutanix nodes. In fact the SATA drives that we use are marketed by Seagate as “Nearline SAS” or NL-SAS. Mainly to differentiate them from the consumer grade SATA drives that are found in cheap laptops. There are hundreds of SAS Vs SATA articles on the web, so I won’t go over the theoretical/historical arguments.
SATA in Hybrid/Tiered Storage
In a Nutanix cluster the “heavy lifting” of IO is mainly done by the SSD’s – leaving the SATA drives to service the few remaining IO’s that miss the SSD tier. Under moderate load, the SATA spindles do pretty well, and since the SATA $/GB is only 60% of SAS. SATA seems like a good choice for mostly-cold data.
Let’s Experiment.
From a performance perspective, I decided to run a few experiments to see just how well SATA performs. In the test, the SATA drives are Nutanix standard drives “ST91000640NS” (Seagate, priced around $150). The comparable SAS drives are the same form-factor (2.5 Inch) “AL13SEB900” (Toshiba, priced at about $250 USD). These drives spin at 10K RPM. Both drives hold around 1TB.
There are three experiments per drive type to reveal the impact of seek-times. This is achieved using the “filesize” parameter of fio – which determines the LBA range to read. One thing to note, is that I use a queue-depth of one. Therefore IOPs can be calculated as simply 1/Response-Time (converted to seconds).
Somewhat intuitively as the working-set (seek) gets larger, the difference between “Real SAS” and “NL-SAS/SATA” gets wider. This is intuitive because with a 1GB working-set, the seek-time is close to zero, and so only the rotational delay (based on RPM) is a factor. In fact the difference in response time is the same as the difference in rotational speed (1:1.3).
Also (just for fun) I used the “random_distribution=zipf” function in fio to test the response time when reading across the entire range of the disk – but with a “hotspot” (zipf) rather than a uniform random read – which is pretty unrealistic.
In the “realistic” case – reading across the entire disk on the SATA drives shipped with Nutanix nodes is capable of 8.5 ms response time at 125 IOPS per spindle.
Conclusion
The performance difference between SAS and SATA is often over-stated. At moderate loads SATA performs well enough for most use-cases. Even when delivering fully random IO over the entirety of the disk – SATA can deliver 8K in less than 15ms. Using a more realistic (not 100% random) access pattern the response time is < 10ms.
For a properly sized Nutanix implementation, the intent is to service most IO from Flash. It’s OK to generate some work on HDD from time-to-time even on SATA.
One of the characteristics of a successful storage system for virtualized environments is that it must handle the IO blender. Put simply, when lots of regular looking workloads are virtualized and presented to the storage, their regularity is lost, and the resulting IO stream starts to look more and more random.
This is very similar to the way that synthesisers work – they take multiple regular sine waves of varying frequencies and add them together to get a much more complex sound.
That’s all pretty awesome for making cool space noises, but not so much when presented to the storage OS. Without the ability to detect regularity, things like caching, pre-fetching and any kind of predictive algorithm break down.
That pre-fetch is never going to happen.
In Nutanix NOS we treat each of these sine waves (workloads) individually, never letting them get mixed together. NDFS knows about vmdk’s or vhdx disks – and so by keeping the regular workloads separate we can still apply all the usual techniques to keep the bits flowing, even at high loads and disparate workload mixes that cause normal storage systems to fall over in a steaming heap of cache misses and metadata chaos.
Although the paper uses examples of both a webserver and a gnutella client, the philosophies are relevant to a large scale distributed filesystem. In the case of NDFS we are serving disk blocks to clients who happen to be virtual machines. One trade-off that is true in both cases is that scalability is traded for low latency in the single-stream case. However at load, the response time is generally better than a system that is designed to low-latency, and then attempted to scale-up to achive high throughput.
At Nutanix we often talk about web-scale architectures, and this paper gives a pretty solid idea of what that might mean in concrete terms.
FWIW., according to google scholar, the paper has been cited 937 times, including Cassandra which is how we store filesystem meta-data in a distributed fashion.
