Monitoring Windows Azure VM Disk Performance
November 13, 2013 1 Comment
Note: The performance data mentioned here is based on individual results during a limited testing period and should NOT be used as an indication of future performance or availability. For more information on Windows Azure storage performance and availability, please refer to the published SLA.
So I’ve had an interesting experience the last few days that I wanted to take a few minutes to share with the interwebs. Namely, monitoring some Windows Azure hosted virtual machines, not at the VM level, but the storage account that held the virtual machine disks.
The scenario I was facing was a customer that was attempting to benchmark the performance and availability of two Linux based virtual machines that were running an Oracle database. Both machines were extra-large VMs with one running 10 disks (1 os, 1 Oracle, 8 data disks) and the other with 16 disks (14 for data). The customer has been running an automated load against both machines and wanted to get a clear idea of how much they may or may not have been saturating the underlying Windows Azure Storage system as well as what could be contributing to the highly variable Oracle IOPS levels they were seeing.
To support this effort, I dug into something I haven’t looked at in depth for quite some time. Windows Azure Storage Analytics (aka Logging and Metrics). Except this time with a focus on what happens at the storage account with regards to the VM disk activity.
Enable Storage Analytics Proactively
Before we go anywhere, I need to stress that if you want to be able to see what’s going on with Azure Storage and your virtual machine, you’ll need to enable this BEFORE a problem occurs. If you haven’t already enabled logging, the only option you have to try and go “back in time” and look at past behavior is to open up a support ticket. So if you plan to do this type of monitoring, please be certain to enable analytics!
For Windows Azure VM disk metrics, we need to enable analytics on the blob storage account. As the link I just shared will let you know, you will need to call the “Set Blob Service Properties” api to set this (or use your favorite Windows Azure storage utility). I happen to use the Azure Management Studio from Redgate and it allows me to set the properties you see in this screen shot:
With this, I tell Azure Storage that I want it to log all blob operations (Read/Write/Delete) and retain that information for up too two days. I also enable metrics and ask it to retain that data for two days as well.
When I enable logging, Azure Storage will log all operations and persist that information into a series of blob files in a special container in the storage account called $logs. The Logging data will be written to blobs in the same storage account I am monitoring in a special container called $logs. Logs will be spread across multiple blob files as discussed in great detail in the MSDN article “About Storage Analytics Logging“. A word of caution, if the storage account is active, logging will produce a LARGE amount of data. In my case, I was seeing a new 150mb log file approximately every 3 minutes. That’s about 70gb per day. In my case, I’ll be storing about 140gb for my 2 days of retention which is only about $6.70 per month. Given the cost of the VM itself, this was inconsequential. But if I had shifted my retention period to a month… this can start to get pricy. Additionally, the storage transactions needed to write the logs to blog storage count against the account limit of 20,000/tps. To help reduce the risk of throttling coming into play to early, the virtual machines I’m monitoring have each been deployed into their own storage account.
The metrics are much more lightweight. These are written to a table and provide a per hour view of the storage account. These are the same values that get surfaced up in the Windows Azure Management portal storage account dashboard. I could easily retain these for a much longer period since it’s only a handful of rows being inserted per hour.
Storage Metrics – hourly summary
Now that we’ve enabled storage analytics and told it to capture the metrics, we can run our test and sit back and look for data to start coming in. After we’ve run testing for several hours, we can then look at the metrics. Metrics get thrown into a series of tables, but since I only care about the blob account, I’m going to look at $MetricsTransactionsBlob. We’ll have multiple rows per hours and can filter based on the type of operation, or get the roll-up across all operations. For general trends, it’s this latter figure I’m most interested in. So I apply a query against the table to get all user operations, “(RowKey eq ‘user;All’)“. The resulting query gives me 1 row per hour that I can look at to help me get a general idea of the performance of the storage account.
You’ll remember that I opted to put my Linux/Oracle virtual machine into its own storage account. So this hour summary gives me a really good, high level overview of the performance of the virtual machine. Key factors I looked at are: Availability (we want to make sure that’s above the storage account 99.9% SLA), Average End to End Latency, and if we have less than 100% availability, what is the count of the number of errors we saw.
I won’t bore you with specific numbers, but over a 24hr period I lowest availability I saw was 99.993% availability and with the most common errors being Server Timeouts, Client Errors, or Network Errors. Seeing these occasionally, as long as the storage account remains above 99.9% availability, should be considered normal ‘noise’. In the transient nature of the cloud, some errors are simply to be expected. We also kept an eye on average end to end latency which during our testing was fairly consistent in the 19-29ms range.
You can learn more about all the data available in these various storage metrics by reviewing ‘Storage Analytics Metrics Table Schema‘ on MSDN.
When we saw numbers that appears “unusual”, we then took the next logical step and inspected the detailed storage logs.
Blob Storage Logs – the devil is in the details
Alright, so things get a bit messier here. First off, the logs are just delimited format files. And while there the metrics can help tell us which period in time we want to look at, depending on the number of storage operations, we may have several logs we need to slog through (In my case, I was getting about 20 150mb log files per hour). So the first step when digging into the logs is to download them. So either write up some code, grab your favorite utility, or perhaps just log into the management portal and download the files for the timeframe you want to take a closer look at. Once that’s done, it’s time for some Excel (yeah, that spreadsheet thing…. Really).
The log files are semi-colon delimited files. As such, the easiest way I found to do ad-hoc inspection of the files is to open them up in a spreadsheet application like Excel. I open up Excel, then do the whole “File -> Open” thing to select the log file I want to look at. I then tell Excel it’s a delimited file with a semi-colon as the delimiter and in a few seconds it will import the file all nice and spreadsheet for me. But before we start doing anything, let’s talk about the log file format. Since the log file doesn’t contain any headers, we either need to know what columns contain the data we want, or add some headers. For the sake of keeping things easy for you (and saving a copy for myself), I created my own Excel file that already has all the log file fields declared in it. So you can just copy and paste from this spreadsheet into your log file once it’s loaded into Excel. For the remainder of this article, I’m going to assume this is what you’ve done.
With our log file headers, we can now start filtering the data. If we’re looking for errors, the first thing we’ll want to do is open up a log file and filter based on “request status”. To do this, select the “Data” tab and click on “filter”. This allows us to click on the various column headings and filter down what we’re looking at. The shot below shows a log that had a couple of errors in it. So I can easily remove the checkbox on “Success” to drill into those specific errors. This is handy if we want to know exactly what happened as the log also contains a “request-id-header” field. With that value, we can open up a support ticket and ask them to dig into the issue more deeply.
Now this is the first real caution I have. Between the metrics and the logs, we can get a really good idea of what types of errors are happening. But this doesn’t mean that every error should be investigated. With cloud computing solutions, there’s a certain amount of “transient” errors that are simply to be expected. It’s only if you see a prolonged, or persistent issue that you’d really want to dig into the errors in any real depth. One key indicator is to look at the logging metrics and keep an eye on the availability. If it falls below 99.9%, that means there may have been an SLA violation for the storage account. In that case, I’d take a look at the logs for that period and see what types of errors we saw. As long as the issue wasn’t caused by a spike in throttling (meaning we overloaded the system), there may be something worth having support look into. But if we’re at 99.999%, with the occasional network failure, timeout, or ‘client other’, we’re likely just seeing the “noise” one would expect from transient errors as the system adjust and compensates for changes to its underlying fabric.
Now since we’re doing benchmarking tests, there’s one other key thing I look at. The number of operations that are occurring on the blobs that are the various disks mounted into our virtual machine. This is another task where Excel can help out, by adding subtotals. Adding subtotals requires column headings so this is the part when you go “thank you Brent for making it so I just need to copy those in”. You’re welcome. J
The field we want to look at in the logs for our subtotal is the “requested-object-key” field. This value is the specific option in the storage account that was being access (aka the blob file or disk). Going again to the Data tab in Excel, we’ll select “subtotal” and complete the dialog box as shown at the left. This will create subtotals by object (disk) and allow us to see the count of operations against that object. So what we have is the operations performed on that disk during the time period covered by the log file. Using that value, we can then get a fairly good approximation of the “transactions per second” that the disk is generating against storage.
So what did we learn?
If you are doing IO benchmarking of the virtual machine (as I was), you may notice something odd. We observed that our Linux/Oracle Vm was reporting IOPS far above what we saw at the Windows Azure Storage level. This is to be expected because Oracle is trying to help buffer requests itself to increase performance. Add in any disk buffering we may have enabled, and the numbers could skew even further. Ultimately though, what we did establish during out testing was that we knew for certain when we were overloading the Windows Azure storage sub-system and contributing to server slowdowns that way. We were also able to observer several small instances where Oracle performance trailed off somewhat and that these were due to isolated incidents where we saw an increase in various errors or in end to end operation latency.
The host result here is that while virtual machine performance is related to the performance of the underlying storage subsystem, there’s no easy 1-to-1 relation between errors in one and issues in the other. Additionally, as you watch these over time, you understand why virtual machine disk performance can vary over time and shouldn’t be compared to the behaviors we’ve come to expect from a physical disk drive. We have also learned what we need to do to help us more affectively monitor Windows Azure storage so that we can proactively take action to address potential customer facing impacts.
I apologize for all the typos and for not going into more depth on this subject. I just wanted to get this all out into before I fell into the fog of my memory. Hopefully you find it useful. So until next time!