Necko/MobileCache/MicroBenchmarks
This Is Work In Progress - Information Is Incomplete
This page describes the cache-related microbenchmarks created so far. Contact BjarneG (bherland@mozilla.com) for a patch which lets you run these benchmark using the "check-one" target of xpcshell tests.
Each benchmark below is described and explained, and we also show output results from selected platforms. Note that results are provided as examples and information only; the benchmarks are meant for evaluating effect of code-changes, i.e. you run the benchmarks without and with your change and compare the results to evaluate the effect of your code.
test_timing_cache.js
This benchmark measures time to load a resource from a server and from the cache, as well as time for call-overhead and clearing the cache.
The approach is to repeat a single load in a loop for some time and report the average time for one iteration. This eliminates timer-granularity issues and should smoothen out random jitter in the system, but the weakness is that IO-caching at any level will influence disk-cache results. Moreover, memory-cache results are likely to be influenced by HW caches and memory-paging in the OS. In short, this benchmark does not implement a very realistic access-pattern for cache-entries, hence its practical value may be limited. It can be useful, though, to study changes unrelated to the speed of the cache media, e.g. code to optimize datastructures or algorithms.
Note that there are two different tests for cache-misses: One which clears the cache in each iteration, and one where the response has a header which prevents it from being cached. (The first is mainly present for completeness - it makes a clearCache() call in each iteration, which is not very likely to happen in real life.) One could also imagine a third approach to emulate cache-misses; suffix the url with a different query-string in order to load a new resource in each iteration. However, this doesn't work because it hits the max number of open connections and times out. (See test below though, where we can control this limit in a different manner.)
All operations are reported for different data-sizes, as well as with no cache enabled, memory-cache enabled only and disk-cache enabled only. (Note that, theoretically, searching for an entry in a cache should take some time, hence we measure also with no cache enabled.)
Some results from a Linux server running 64-bit Ubuntu 10.4
Total time #runs avg time/run === Set datasize 50 bytes Pure overhead (no-cache) 10015 14902 0.67 Overhead+ClearCache (no-cache) 10016 11703 0.86 Load (no cache) 10041 362 27.74 Pure overhead (mem-cache) 10016 14799 0.68 Overhead+ClearCache (mem-cache) 10016 11663 0.86 Cache hit (mem-cache) 10018 4985 2.01 Cache miss/clrcache (mem-cache) 10038 347 28.93 Cache miss/nocache (mem-cache) 10046 365 27.52 Pure overhead (disk-cache) 10015 14832 0.68 Overhead+ClearCache (disk-cache) 10015 6674 1.50 Cache hit (disk-cache) 10017 4973 2.01 Cache miss/clrcache (disk-cache) 10019 343 29.21 Cache miss/nocache (disk-cache) 10029 363 27.63 === Set datasize 1024 bytes Load (no cache) 10020 351 28.55 Cache hit (mem-cache) 10019 3829 2.62 Cache miss/nocache (mem-cache) 10037 349 28.76 Cache hit (disk-cache) 10018 3800 2.64 Cache miss/nocache (disk-cache) 10018 344 29.12 === Set datasize 51200 bytes Load (no cache) 10093 108 93.45 Cache hit (mem-cache) 10053 324 31.03 Cache miss/nocache (mem-cache) 10051 106 94.82 Cache hit (disk-cache) 10060 325 30.95 Cache miss/nocache (disk-cache) 10039 107 93.82 === Set datasize 524288 bytes Load (no cache) 10557 15 703.80 Cache hit (mem-cache) 10404 33 315.27 Cache miss/nocache (mem-cache) 10577 15 705.13 Cache hit (disk-cache) 10354 33 313.76 Cache miss/nocache (disk-cache) 10604 15 706.93
The important numbers are found in the rightmost column. Note that the first block (datasize 50 bytes) includes more results than the other blocks. "Pure Overhead" measures the call-overhead from JavaScript, and "Overhead+ClearCache" measures the time to Clear the cache from JavaScript. These numbers are unrelated to the size of the data loaded, hence reported only once.
Results above indicate that reading from the disk-cache is about as fast as reading from memory-cache, which probably is due to efficient IO-caching on this platform. Note also that cache-hits for small entries are relatively much faster than cache-hits for larger entries (i.e. for small entries there is a factor 10 whereas for large entries we see a factor 2-3).
There is no conclusive evidence that searching a disk or memory cache takes any significant amount of time. However, the caches in this test are pretty much empty so we should create a separate test for this, making sure the cache contains some number of entries before we measure. (See #Wanted_tests below)
test_timing_cache_2.js
This benchmark loads a number of resources with different cache-keys and then waits for all cache-io to finish. Then it loads all resources again (retrieving them from cache this time). Both sequences are measured and reported for various entry-sizes and cache-configurations.
This access-pattern is slightly more realistic than in the benchmark described above; a page is likely to load some number of resources from the (same) server, and if the user visits the page second time, those resources are loaded again (from cache this time). The benchmark aims at emulating this pattern in a simple way.
Some results from a Linux server running 64-bit Ubuntu 10.4
Number of entries: 100 Max #connections=256 / max per server=256
Total time #runs avg time/run
Setting datasize 50 bytes
Load from server (no cache) 2324 101 23.01
Load from server (mem-cache) 2346 101 23.23
Load from cache (mem-cache) 208 101 2.06
Load from server (disk-cache) 2334 101 23.11
Load from cache (disk-cache) 180 101 1.78
Setting datasize 1024 bytes
Load from server (no cache) 2419 101 23.95
Load from server (mem-cache) 2409 101 23.85
Load from cache (mem-cache) 234 101 2.32
Load from server (disk-cache) 2434 101 24.10
Load from cache (disk-cache) 247 101 2.45
Setting datasize 51200 bytes
Load from server (no cache) 9056 101 89.66
Load from server (mem-cache) 9392 101 92.99
Load from cache (mem-cache) 3100 101 30.69
Load from server (disk-cache) 9465 101 93.71
Load from cache (disk-cache) 3077 101 30.47
Setting datasize 524288 bytes
Load from server (no cache) 71864 101 711.52
Load from server (mem-cache) 72233 101 715.18
Load from cache (mem-cache) 17015 101 168.47
Load from server (disk-cache) 84842 101 840.02
Load from cache (disk-cache) 29878 101 295.82
The interesting numbers are found in the rightmost column of the output.
Results above suggest that loading entries up to a certain size from the server (the xpcshell-handler in this case) is rather consistent regardless of whether we use no cache, mem-cache or a disk-cache. Also, when loading entries up to a certain size from cache, the disk-cache seems to be as fast as the memory-cache. This IMO indicates that IO on this platform is pretty efficient (up to a certain limit) because we would always expect to spend more time when using the disk-cache than when using mem-cache. Observe that we need 512K-sized resources to see a clear difference between using disk and memory. (The exact limit here is not identified, but it may not very interesting either since it is probably platform-specific).
Also observe that when reading from the cache, small entries load relatively much faster than larger entries. (Memory-page size?)
Wanted tests
Unordered and informal list of tests we may also want to create
- measure time to search a cache
- mem- and disk-only, but also the combination of them I believe. Make sure we search the cache without finding the entry.
- identify entry-size where disk-cache is clearly slower than mem-cache
- probably OS- or device-specific but it may be worth finding in case we can learn something general from it
- driver which can take a list of resources and measure time to load them all
- we can extract this info from e.g. Talos or by using telemetry, and it should probably be hierarchical (some resources causes other to load)