I’ve always had a focus on storage throughout my career. I’ve managed large enterprise vSANs with FC switches, commercial NAS filers, deployed iSCSI over ethernet, and managed ESX with both FC and NFS backends. I’ve been entrusted to build very large storage servers, up to 32U, with Linux and off the shelf components. Needless to say, I feel comfortable claiming that I know a little more than the average systems guy about storage, and particularly how Linux handles I/O, so when I turned my attention to benchmarking virtual machine disk performance, I found some interesting behaviors that most who seek to measure such things should probably be aware of, at least to interpret results, if they can’t otherwise be compensated for.
One of the primary things is how the Linux caching mechanisms can throw a wrench in things if you don’t think through what you’re doing. One needs to be aware of which caches are in effect during each test. For example, it’s common to test with datasets larger than the system’s memory in order to stretch the system beyond its ability to cache, however, consider a 4GB virtual guest on a physical server with 32GB RAM. Usually the guest systems are run with at least write-through cache from the host’s perspective (speaking in general terms, this can obviously be controlled by the end user on at least some virtualization platforms), so while the experimenter might think that using an 8GB dataset will be sufficient on the guest, or that issuing a drop_caches request between tests on the guest will suffice, this dataset is likely to be saved in its entirety in the host’s read cache as it goes to underlying storage, artificially boosting the results. Similarly, performing a write test on the guest and comparing it to the same write test on the host is almost certainly going to give the host an unfair advantage if the experimenter doesn’t take into account the increase in dirty memory available on the host, usually specified in percent of physical memory.
On top of that, there’s the complexity of testing X number of virtual machines and forming a summation of how they all perform simultaneously on a physical host. There are some pretty standard methods defined for doing this, such as putting some sort of load on each guest, and then benchmarking one while the others are running their dummy loads, but again, one must be careful, particularly with the dummy loads, that they’re not just looping tests that are small enough to cache, unless, of course, that’s the real-world behavior of the application, which brings me to my point.
It’s kind of a complex beast, trying to get meaningful results, and especially to share them with others who may have different expectations. One has to determine a goal in disk benchmarking, and it’s usually one of two things; the testing of raw disk performance or an attempt to measure the real-world performance of an application or given I/O pattern. The former would involve disabling any and all caches, while the latter would strive to utilize the caches how they normally would be. The challenge in all of this, as mentioned, is that some people will value one set, while others will value the other. Raw disk performance will tell you a lot about just how good the setup is, for example whether one should go with that raid6 setup or do raid50 instead, on the other hand, does it really matter how well the disks perform without caches, don’t we want to know how it’s actually going to run?
No matter how it’s done, the most important thing of all is to frame your data properly. “This was the goal or purpose, these are the tests, this is the setup, here are the results”. I’ve been running some tests that I’ll share shortly, but I wanted to get some of these cosiderations down, as I’ve rarely heard anyone speak of them while reading through the benchmarks of others, which frankly, has made much of the data I’ve seen surrounding vm performance largely useless.
Finally, lest this post be all rambling and not provide anything of concrete usefulness to individuals out there, the following are some mechanisms for controlling Linux caching.
Flush caches (page, dentries, inodes): ‘echo 3 > /proc/sys/vm/drop_caches’
The above won’t do anything for dirty memory, which can be cleaned up with a ‘sync’, however, this won’t have much bearing on the write test you run afterward, you’ll need to know a little more about how dirty memory works. It would be naive to compare a system with 32G of memory, 3.2 of which can absorb pending writes, with a 4G system that only has 400M with which to cache writes.
In particular, two values are of importance: /proc/sys/vm/dirty_ratio and /proc/sys/vm/dirty_background_ratio. These two numbers are specified as percentages. dirty_background_ratio tells you how big your dirty memory can get before pdflush kicks in and starts writing it out to disk. dirty_ratio is always higher (the code actually rewrites dirty_background_ratio to half of this if dirty_ratio < dirty_background_ratio), and is the point where applications skip dirty memory and are forced to write direct to disk. Usually this means that pdflush isn’t keeping up with your writes, and the system is potentially in trouble, but could also just mean that you’ve set it very low because you don’t want to cache writes. For example, you may want to do this if you know you’ll be doing monster writes for extended periods, no sense in bloating up some huge amount of dirty memory only to have the processes forced to write sync AND contend with pdflush threads trying to do writeback. On the flip side, increasing these values can give you a nice cache to absorb large, intermittent writes.
Both of these have time based counterparts, dirty_expire_centisecs and dirty_writeback_centisecs, such that pdflush will kick in and start doing writeback by age regardless of how much is there. For example, it might do writeback at 500MB OR when data in dirty memory has been around for longer than 15 seconds. Newer kernels also allow an alternative specification of an actual number, rather than percent, in dirty_bytes and dirty_background_bytes.
There are quite a few more things I could share, but I think I’ll leave with just one more: /proc/sys/vm/vfs_cache_pressure. Usually this is set at 100 by default. Increasing this number will cause the system to tend to clean up/minimize directory and inode read caches (the stuff that’s cleaned up by drop_caches), decreasing the number will cause it to horde more.
Stay tuned for some benchmarks of KVM virtio and IDE with no cache, writethrough, and writeback, compared to VMware ESX paravirtualized disks.