ACC / YOTTASERVE are a solutions company specializing in state of the art ‘boot from clustered SAN’ hosting environments. These hosting environments enable users to achieve complete separation of operating system and server hardware, so that it’s possible to ‘swap out’ failed hardware and boot the OS from working hardware all remotely. The diskless server environment also uses much less physical space, and far less power than traditional hosting methods.
SANs are all about I/O. There are many I/O based benchmarks out there – the most established is I/OMeter. We have run exhaustive testing diskless volumes on the SAN, IOMeter shows the SAN significantly outperforms typical hard disk setups by a large margin. However, IOMeter is a benchmark that simply tests the SAN as an I/O subsystem, and not really as a diskless server, running real life applications. Customers demand real life comparative test data to believe that running diskless servers over gigabit on a clustered SAN, would offer the same, or better performance than servers that have locally connected hard disks. It’s a concept people just can’t get their head around. They need concrete proof.
We configured a typical Windows 2003 server setup, running Coldfusion MX, MS SQL Server, and a Database Driven content management system, running a copy of the Yottaserve website.
NeoLoad was used to generate a rapid ramp of website users to the target server website, using a set scenario. We were easily able to create NeoLoad WMI monitors to log the target servers critical metrics, such as disk, CPU, memory, database hits, application server requests.
We then migrated the servers hard disk onto a volume on the clustered SAN and booted the same server to this volume, which was connected via a gigabit switch. The same load tests were run again.
NeoLoads excellent reporting system enabled us to run multiple tests, and keep track of all of them, instantly providing test by test comparisons. The results were there in front of us and plain to see. For the first time we could now prove to interested customers that for running real life applications, the SAN delivers similar or better performance than locally connected hard disks.
Our diskless servers connect to SANs via iSCSI – a storage protocol that runs inside the IP network protocol, and connects to servers over standard Gigabit ethernet. Normally one would need to buy a hardware card – a ‘Host Bus Adapter’ (HBA) to be able to connect a server to an iSCSI disk over a network. However, Microsoft have a software HBA – freely available – the Microsoft iSCSI Initiator. The problem is that, as this is software, and not hardware, the extra ‘work’ needed to be done to process the iSCSI packets is done by the server CPU, and not a hardware card. So in this way iSCSI does impact the performance of your server when under load. Our newest servers however are equipped with special ‘I/O’ acceleration hardware , embedded on the motherboard chipset – a setting which is enabled in BIOS. So you get the benefits of I/O acceleration, without having to add an extra card to the server. However, it was always difficult for us to know how well this feature was working – if at all!
So we ran the load test with the setting enabled and then disabled, and connected NeoLoad to the CPU, so we could track the % CPU utilization. If the I/O acceleration (TOE -TCP Offload Engine) was working, we should see a reduction in CPU on the test when the feature was enabled. We didn’t. After some further investigations, it turned out there was a ambiguous chipset driver problem, and after it was sorted out, the tests were re-run, and our modest 5% – small but yet significant – reduction in CPU utilization was clearly evident!
Résultats du test et valeur ajoutée
Application testing, server hardware testing is quite a laborious and boring process. What makes this task so much more un-appealing is simply the admin of having to fire up excel and keep all results typed down. If you want to run comparisons, that’s lots of excel work and quite frankly there are better things I can be doing with my time. I have found that NeoLoad drastically cuts down the amount of ‘admin’ needed to log test results, and report on them. This means I can focus on analyzing at the data, rather than logging it. This also means we can focus on running more tests, which will only result in us delivering better products to our customers.
So whether I’m testing hardware or software, NeoLoad puts load is where is should be: on my test kit, and not on me!