Necko/MobileCache/MicroBenchmarks
This Is Work In Progress - Information Is Incomplete
This page describes the cache-related microbenchmarks. Contact BjarneG (bherland@mozilla.com) for further information and/or the benchmarks themselves.
Purpose
The primary purpose of the microbenchmarks is to be tools for programmers evaluate the effect of code-changes on cache-performance. I.e. you run benchmarks, record results, apply code-changes, then run the benchmarks again. The goal is that comparing two results should give a reasonable idea how your changes affects cache-performance.
Implications of the above include
- we need to be able to reproduce results rather consistently, i.e. we cannot have (too) different results from different runs
- we might have to process raw results somehow; perhaps use some statistical properties
- at the very least we must quickly be able to see variance/deviation in a dataset
- what we measure (e.g. which units we use) is not really important - we just need something comparable
A secondary purpose of these microbenchmarks is to uncover performance-problems on new platforms. For example, enabling the disk-cache is rumored to hurt performance on mobile devices, based on running TP4 on one or two devices with and without disk cache enabled. With a reasonable set of benchmarks we can pinpoint more precisely what the problem is, which devices it affects, and do something about it.
Implications of the above include
- we must somehow be able to identify a bad test-result (i,e, we cannot just report reproducible, abstract numbers - the results must be somehow understandable)
- we need many measurements
- search time vs cache sizes vs full/empty cache
- read-time vs cache-type vs large/small data
- write-time vs cache-type vs large/small data (writing is off main-thread, but will probably still impact performance on devices?)
- creating, clearing, adjusting size, dooming entries
- with many and detailed measurements we most likely also need a way to depict results graphically
Telemetry vs microbenchmarks
There has been some discussion about using telemetry instead of xpcshell-based microbenchmarks. Current thinking is that they are complementary: Telemetry is real-life browsing patterns on real-life platforms and environments, whereas microbenchmarks are artificial browsing patterns running on a machine in a lab-environment. It is impractical to experiment with code-changes using telemetry to measure the effect - a benchmark in the lab is much more practical for this. On the other hand, telemetry is the (only?) way to ensure that improvements also are valid in real-life. Moreover, with telemetry it is very difficult (if not impossible) to get the context of your measurements. For example, suppose you measure the time to evict a cache-entry from disk-cache; for this measurement to make sense you also need to know the number of entries in the cache and the total size of the cache. This context-information is hard to get using telemetry alone.
We plan to use telemetry as described below.
Identify areas of interest
Telemetry will provide lots of data and a really important job is to read and analyze this. It is expected that we will see unexpected patterns from telemetry, and such patterns should trigger microbenchmarks to investigate specific areas.
Tune parameters to make microbenchmarks more realistic
Original conditions in the lab are bound to be different from what users see in real life. However, in the lab we can control a number of parameters; by knowing realistic values for these parameters we can run tests in the lab under similar conditions as users, making our tests more realistic.
Examples of this include
- Bandwidth: What's the typical bandwidth(s) seen by a user? In the lab we normally load resources from the local host...
- Latency/RTT: How long does it typically take from sending a request to the server before the response starts to appear?
- Cache-sizes: What's the typical cache-size out there?
Real-life verification of results from the lab
One way to define network-performance on a given platform is as the product of two factors: The browsing pattern (i.e. which urls are loaded in which sequence), and what exactly is measured. As discussed above, microbenchmarks and telemetry are inherently different with respect to the browsing pattern, but we can do our best to align telemetry and microbenchmarks wrt the second factor. Put in a different way: We should try to use the same code in telemetry and microbenchmarks to capture data, and we should ensure we interpret this data in the same way.
The major benefit of this is to have telemetry give us real-life verification after using synthetic, isolated and focused benchmarks in the lab. I.e. we can use synthetic test-patterns in the lab to identify and qualify code-changes, then after landing code-changes we should be able to see predictable effects of these changes via telemetry. If we measure performance differently in microbenchmarks and telemetry we may quickly end up "comparing apples and oranges", confusing ourselves.
Below is a pro/con list for using telemetry-code vs JS time-functions to capture data for microbenchmarks - feel free to add and comment.
| Time-function in JavaScript | Telemetry timestamps compiled into the browser | |
|---|---|---|
| Pros |
|
|
| Cons |
|
|
The benchmarks
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_stresstest_cache_with_timing.js
This test loads a number of different resources with uniform size. The resources are loaded sequentially in a tight loop and the test syncs with cache-io thread after finishing the loop. Each loop is executed for five different cache-configurations/setups:
- no cache enabled, loading all urls from server, not writing to cache (clearing cache before this step)
- memory-cache enabled, loading all urls from server, write them to cache (clearing cache before this step)
- memory-cache enabled, read all urls from cache (resources cached in previous step)
- disk-cache enabled, loading all urls from server, write them to cache (clearing cache before this step)
- disk-cache enabled, reading all urls from cache (resources cached in previous step)
This whole procedure is repeated over a given list of datasizes, clearing the cache between each repetition.
Time is measured in two ways: 1) for each individual load by by using the nsITimingChannel interface of the channel, and 2) by measuring run-time of the complete loop from the test itself.
Informally one can say that this test loads a large number of uniformly sized resources at the same time, somewhat similar to what happens when starting up firefox with lots of open tabs.
Generated datafiles
This test writes data into a number of files. Note that all files are appended, meaning that if they exist prior to the run, you will have strange data...
- summary.dat
- Format is multi-data with one datablock for each datasize run by the test. Each datablock has a header describing the datasize, then one row for each of the five configurations summing up the results from the loop. Datapoints are
- description of the configuration
- minimum value from the loop
- maximum value from the loop
- calculated mean over the results from the loop
- standard deviation over all data from the loop
- number of iterations in the loop
- time measured by the test using Data.now() for the loop
- time measured by the test using Data.now() for the loop, including syncing with the cache-io thread at the end
- telemetry-<bytes>.dat
- One file for each datasize run by the test. The format is multi-data with one datablock for each configuration. Each row represents one resource-load in actual order. The datapoints in each row are
- time from create to asyncOpen
- time from asyncOpen to cacheEntryAvailable
- time from cacheEntryAvailable to cacheReadStart
- time from cacheReadStart to cacheReadEnd
alternatively (if loaded from channel, not from cache)
- time from create to asyncOpen
- time from asyncOpen to cacheEntryAvailable
- time from cacheEntryAvailable to responseStart
- time from responseStart to responseEnd
I.e. referring to the Telemetry-data accumulated by nsLoadGroup, (2+3) corresponds to Telemetry::HTTP_PAGE_OPEN_TO_FIRST_FROM_CACHE alternatively Telemetry::HTTP_PAGE_OPEN_TO_FIRST_RECEIVED, and (2+3+4) corresponds to Telemetry::HTTP_PAGE_COMPLETE_LOAD.
Files to work with GnuPlot (requires Gnuplot >= 4.4)
- settings.gnu
- setup for the histogram, title of plot etc. Edit this to describe your platform.
- plot-summary.gnu
- Plots mean-values from summary.dat in one histogram, grouped by first column. Standard deviation in the dataset is indicated as an errorline. This is the recommended starting-point for studying results from a test-run.
- plot-minmax-summary.gnu
- Like above but indicates min and max values from the dataset instead of stddev.
- plot-summary-per-size.gnu and plot-minmax-summary-per-size.gnu
- Plots summaries like the two plots above, but makes one plot for each datasize to avoid scaling-issues in the plot.
- plot-summary-from-javascript.gnu
- This is WIP. Plots summaries of average loading-times seen from JS.
- plot-telemetry.gnu
- You need to copy the file "telemetry-<size>.dat" to "telemetry.dat" for the size you want to study. Plots detailed information for each resource loaded, one plot for each configuration. The format used is a rowstacked histogram where each column shows total time from creating the channel to loading has finished. Note that columns show time for each load separately (all times starting at 0) as opposed to a real time-line for the dataset.
test_timing_cache.js (Obsolete)
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
Memory-cache capacity: 51200 Kb / Disk-cache capacity: 1048576 Kb
Max #connections=256 / max per server=256
Total time #runs avg time/run
=== Set datasize 50 bytes
Calloverhead (no-cache) 10016 14594 0.69
Calloverhead+ClearCache (no-cache) 10015 11647 0.86
CacheMiss (nocache) (no cache) 10047 368 27.30
Calloverhead (mem-cache) 10015 14532 0.69
Calloverhead+ClearCache (mem-cache) 10016 11616 0.86
CacheMiss (clrcache) (mem-cache) 10018 360 27.83
CacheMiss (nocache) (mem-cache) 10032 367 27.34
CacheHit (mem-cache) 10017 4674 2.14
Calloverhead (disk-cache) 10015 13954 0.72
Calloverhead+ClearCache (disk-cache) 10017 6723 1.49
CacheMiss (clrcache) (disk-cache) 10033 346 29.00
CacheMiss (nocache) (disk-cache) 10032 357 28.10
CacheHit (disk-cache) 10017 4709 2.13
=== Set datasize 1024 bytes
CacheMiss (nocache) (no cache) 10020 339 29.56
CacheMiss (nocache) (mem-cache) 10037 351 28.60
CacheHit (mem-cache) 10019 3634 2.76
CacheMiss (nocache) (disk-cache) 10048 345 29.12
CacheHit (disk-cache) 10018 3535 2.83
=== Set datasize 51200 bytes
CacheMiss (nocache) (no cache) 10095 103 98.01
CacheMiss (nocache) (mem-cache) 10023 104 96.38
CacheHit (mem-cache) 10070 306 32.91
CacheMiss (nocache) (disk-cache) 10110 104 97.21
CacheHit (disk-cache) 10054 301 33.40
=== Set datasize 524288 bytes
CacheMiss (nocache) (no cache) 10277 14 734.07
CacheMiss (nocache) (mem-cache) 10312 14 736.57
CacheHit (mem-cache) 10443 31 336.87
CacheMiss (nocache) (disk-cache) 10283 14 734.50
CacheHit (disk-cache) 10451 31 337.13
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)
Some raw results from my Nexus S. I find it weird that this device outperforms the Linux server when the data-size becomes larger... any insight is appreciated
Memory-cache capacity: 51200 Kb / Disk-cache capacity: 1048576 Kb
Max #connections=256 / max per server=256
Total time #runs avg time/run
=== Set datasize 50 bytes
Calloverhead (no-cache) 10016 10878 0.92
Calloverhead+ClearCache (no-cache) 10016 9482 1.06
CacheMiss (nocache) (no cache) 10020 321 31.21
Calloverhead (mem-cache) 10015 10975 0.91
Calloverhead+ClearCache (mem-cache) 10016 8937 1.12
CacheMiss (clrcache) (mem-cache) 10036 371 27.05
CacheMiss (nocache) (mem-cache) 10021 348 28.80
CacheHit (mem-cache) 10016 4639 2.16
Calloverhead (disk-cache) 10016 11426 0.88
Calloverhead+ClearCache (disk-cache) 10043 213 47.15
CacheMiss (clrcache) (disk-cache) 10034 117 85.76
CacheMiss (nocache) (disk-cache) 10019 347 28.87
CacheHit (disk-cache) 10017 4206 2.38
=== Set datasize 1024 bytes
CacheMiss (nocache) (no cache) 10049 332 30.27
CacheMiss (nocache) (mem-cache) 10024 339 29.57
CacheHit (mem-cache) 10017 3501 2.86
CacheMiss (nocache) (disk-cache) 10031 327 30.68
CacheHit (disk-cache) 10017 3308 3.03
=== Set datasize 51200 bytes
CacheMiss (nocache) (no cache) 10064 119 84.57
CacheMiss (nocache) (mem-cache) 10074 119 84.66
CacheHit (mem-cache) 10049 396 25.38
CacheMiss (nocache) (disk-cache) 10078 116 86.88
CacheHit (disk-cache) 10031 426 23.55
=== Set datasize 524288 bytes
CacheMiss (nocache) (no cache) 10572 19 556.42
CacheMiss (nocache) (mem-cache) 10052 18 558.44
CacheHit (mem-cache) 10274 45 228.31
CacheMiss (nocache) (disk-cache) 10167 18 564.83
CacheHit (disk-cache) 10177 43 236.67
This is one way to represent the results
| No cache | Memory-Cache | Disk-Cache | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 50 | 1024 | 51200 | 524288 | 50 | 1024 | 51200 | 524288 | 50 | 1024 | 51200 | 524288 | ||
| Ubuntu 64bit server | Clear cache | 0.86 | 0.86 | 1.49 | |||||||||
| Miss | 27.30 | 29.56 | 98.01 | 734.07 | 27.34 | 28.60 | 96.38 | 736.57 | 28.10 | 29.12 | 97.21 | 734.50 | |
| Hit | n/a | 2.14 | 2.76 | 32.91 | 336.87 | 2.13 | 2.83 | 33.40 | 337.13 | ||||
| Nexus S | Clear cache | 1.06 | 1.12 | 47.15 | |||||||||
| Miss | 31.21 | 30.27 | 84.57 | 556.42 | 28.80 | 29.57 | 84.66 | 558.44 | 28.87 | 30.68 | 86.88 | 564.83 | |
| Hit | n/a | 2.16 | 2.86 | 25.38 | 228.31 | 2.38 | 3.03 | 23.55 | 236.67 | ||||
| Samsung Galaxy S (from Geoff) | Clear cache | 1.89 | 1.87 | 1339.75 | |||||||||
| Miss | 31.33 | 27.48 | 74.04 | 524.20 | 27.27 | 28.32 | 75.05 | 529.32 | 28.04 | 28.43 | 93.92 | 612.59 | |
| Hit | n/a | 2.67 | 3.04 | 21.67 | 209.69 | 3.03 | 3.37 | 24.37 | 218.61 | ||||
and this is another representation. I prefer the latter since it allows me to easier study the effect of different cache-configurations on a particular testcase
| 50 | 1024 | 51200 | 524288 | ||
|---|---|---|---|---|---|
| Ubuntu 64bit server | Miss (no-cache) | 27.30 | 29.56 | 98.01 | 734.07 |
| (mem-cache) | 27.34 | 28.60 | 96.38 | 736.57 | |
| (disk-cache) | 28.10 | 29.12 | 97.21 | 734.50 | |
| Hit (mem-cache) | 2.14 | 2.76 | 32.91 | 336.87 | |
| (disk-cache) | 2.13 | 2.83 | 33.40 | 337.13 | |
| Nexus S | Miss (no-cache) | 31.21 | 30.27 | 84.57 | 556.42 |
| (mem-cache) | 28.80 | 29.57 | 84.66 | 558.44 | |
| (disk-cache) | 28.87 | 30.68 | 86.88 | 564.83 | |
| Hit (mem-cache) | 2.16 | 2.86 | 25.38 | 228.31 | |
| (disk-cache) | 2.38 | 3.03 | 23.55 | 236.67 | |
| Samsung Galaxy S (from Geoff) | Miss (no-cache) | 31.33 | 27.48 | 74.04 | 524.20 |
| (mem-cache) | 27.27 | 28.32 | 75.05 | 529.32 | |
| (disk-cache) | 28.04 | 28.43 | 93.92 | 612.59 | |
| Hit (mem-cache) | 2.67 | 3.04 | 21.67 | 209.69 | |
| (disk-cache) | 3.03 | 3.37 | 24.37 | 218.61 | |
test_timing_cache_2.js (Obsolete)
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
Memory-cache capacity: 51200 Kb / Disk-cache capacity: 1048576 Kb
Max #connections=256 / max per server=256
Total time #runs avg time/run
Setting datasize 50 bytes
CacheMiss (nocache) (no cache) 2803 101 27.75
CacheMiss (q-url) (mem-cache) 2767 101 27.40
CacheHit (mem-cache) 205 101 2.03
CacheMiss (q-url) (disk-cache) 2783 101 27.55
CacheHit (disk-cache) 186 101 1.84
Setting datasize 1024 bytes
CacheMiss (nocache) (no cache) 2884 101 28.55
CacheMiss (q-url) (mem-cache) 2570 101 25.45
CacheHit (mem-cache) 269 101 2.66
CacheMiss (q-url) (disk-cache) 2910 101 28.81
CacheHit (disk-cache) 271 101 2.68
Setting datasize 51200 bytes
CacheMiss (nocache) (no cache) 9546 101 94.51
CacheMiss (q-url) (mem-cache) 9536 101 94.42
CacheHit (mem-cache) 3274 101 32.42
CacheMiss (q-url) (disk-cache) 9543 101 94.49
CacheHit (disk-cache) 3274 101 32.42
Setting datasize 524288 bytes
CacheMiss (nocache) (no cache) 75942 101 751.90
CacheMiss (q-url) (mem-cache) 75840 101 750.89
CacheHit (mem-cache) 27799 101 275.24
CacheMiss (q-url) (disk-cache) 79318 101 785.33
CacheHit (disk-cache) 31333 101 310.23
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?)
Results from my Nexus S. Like above, the performance on large data is surprisingly good compared to the Linux server... any insight?
Memory-cache capacity: 51200 Kb / Disk-cache capacity: 1048576 Kb
Max #connections=256 / max per server=256
Total time #runs avg time/run
Setting datasize 50 bytes
CacheMiss (nocache) (no cache) 2724 101 26.97
CacheMiss (q-url) (mem-cache) 2249 101 22.27
CacheHit (mem-cache) 126 101 1.25
CacheMiss (q-url) (disk-cache) 2314 101 22.91
CacheHit (disk-cache) 146 101 1.45
Setting datasize 1024 bytes
CacheMiss (nocache) (no cache) 2190 101 21.68
CacheMiss (q-url) (mem-cache) 2206 101 21.84
CacheHit (mem-cache) 177 101 1.75
CacheMiss (q-url) (disk-cache) 2382 101 23.58
CacheHit (disk-cache) 189 101 1.87
Setting datasize 51200 bytes
CacheMiss (nocache) (no cache) 7388 101 73.15
CacheMiss (q-url) (mem-cache) 7582 101 75.07
CacheHit (mem-cache) 2205 101 21.83
CacheMiss (q-url) (disk-cache) 7749 101 76.72
CacheHit (disk-cache) 2226 101 22.04
Setting datasize 524288 bytes
CacheMiss (nocache) (no cache) 57251 101 566.84
CacheMiss (q-url) (mem-cache) 57939 101 573.65
CacheHit (mem-cache) 22888 101 226.61
CacheMiss (q-url) (disk-cache) 62850 101 622.28
CacheHit (disk-cache) 22626 101 224.02
Wanted tests
Unordered and informal list of tests we may also want to create
- variation of the other tests where we start out with a full cache
- i.e. investigate if there is a significant impact of evicting entries while we use the cache
- 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 [Bjarne works on this]
- 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)