1page.title=Low RAM 2@jd:body 3 4<!-- 5 Copyright 2013 The Android Open Source Project 6 7 Licensed under the Apache License, Version 2.0 (the "License"); 8 you may not use this file except in compliance with the License. 9 You may obtain a copy of the License at 10 11 http://www.apache.org/licenses/LICENSE-2.0 12 13 Unless required by applicable law or agreed to in writing, software 14 distributed under the License is distributed on an "AS IS" BASIS, 15 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 16 See the License for the specific language governing permissions and 17 limitations under the License. 18--> 19<div id="qv-wrapper"> 20 <div id="qv"> 21 <h2>In this document</h2> 22 <ol id="auto-toc"> 23 </ol> 24 </div> 25</div> 26 27<h2 id="intro">Introduction</h2> 28 29<p>Android now supports devices with 512MB of RAM. This documentation is intended 30to help OEMs optimize and configure Android 4.4 for low-memory devices. Several 31of these optimizations are generic enough that they can be applied to previous 32releases as well.</p> 33 34<h2 id="optimizations">Android 4.4 platform optimizations</h2> 35 36<h3 id="opt-mgmt">Improved memory management</h3> 37<ul> 38<li>Validated memory-saving kernel configurations: Kernel Same-page Merging 39(KSM), and Swap to ZRAM.</li> 40<li>Kill cached processes if about to be uncached and too large.</li> 41<li>Don’t allow large services to put themselves back into A Services (so they 42can’t cause the launcher to be killed).</li> 43<li>Kill processes (even ordinarily unkillable ones such as the current IME) 44that get too large in idle maintenance.</li> 45<li>Serialize the launch of background services.</li> 46<li>Tuned memory use of low-RAM devices: tighter out-of-memory (OOM) adjustment 47levels, smaller graphics caches, etc.</li> 48</ul> 49 50<h3 id="opt-mem">Reduced system memory</h3> 51<ul> 52<li>Trimmed system_server and SystemUI processes (saved several MBs).</li> 53<li>Preload dex caches in Dalvik (saved several MBs).</li> 54<li>Validated JIT-off option (saves up to 1.5MB per process).</li> 55<li>Reduced per-process font cache overhead.</li> 56<li>Introduced ArrayMap/ArraySet and used extensively in framework as a 57lighter-footprint replacement for HashMap/HashSet.</li> 58</ul> 59 60<h3 id="opt-proc">Procstats</h3> 61<p> 62Added a new Developer Option to show memory state and application memory usage 63ranked by how often they run and amount of memory consumed. 64</p> 65 66<h3 id="opt-api">API</h3> 67<p> 68Added a new ActivityManager.isLowRamDevice() to allow applications to detect 69when running on low memory devices and choose to disable large-RAM features. 70</p> 71 72<h3 id="opt-track">Memory tracking</h3> 73<p> 74New memtrack HAL to track graphics memory allocations, additional information 75in dumpsys meminfo, clarified summaries in meminfo (for example reported free 76RAM includes RAM of cached processes, so that OEMs don’t try to optimize the 77wrong thing). 78</p> 79 80<h2 id="build-time">Build-time configuration</h2> 81<h3 id="flag">Enable Low Ram Device flag</h3> 82<p>We are introducing a new API called <code>ActivityManager.isLowRamDevice()</code> for applications to determine if they should turn off specific memory-intensive 83 features that work poorly on low-memory devices.</p> 84<p>For 512MB devices, this API is expected to return: "true" It can be enabled by 85 the following system property in the device makefile.<br/> 86<code>PRODUCT_PROPERTY_OVERRIDES += ro.config.low_ram=true</code></p> 87 88<h3 id="jit">Disable JIT</h3> 89 90 <p>System-wide JIT memory usage is dependent on the number of applications 91 running and the code footprint of those applications. The JIT establishes a 92 maximum translated code cache size and touches the pages within it as needed. 93 JIT costs somewhere between 3M and 6M across a typical running system.<br/> 94 <br/> 95 The large apps tend to max out the code cache fairly quickly (which by default 96 has been 1M). On average, JIT cache usage runs somewhere between 100K and 200K 97 bytes per app. Reducing the max size of the cache can help somewhat with 98 memory usage, but if set too low will send the JIT into a thrashing mode. For 99the really low-memory devices, we recommend the JIT be disabled entirely.<code> 100</code></p> 101 102<p>This can be achieved by adding the following line to the product makefile:<br/> 103<code>PRODUCT_PROPERTY_OVERRIDES += dalvik.vm.jit.codecachesize=0</code></p> 104<h3 id="launcher">Launcher Configs</h3> 105 106 107 <p>Ensure the default wallpaper setup on launcher is <strong>not</strong> 108using live-wallpaper. Low-memory devices should not pre-install any live wallpapers. </p> 109 110 111<h2 id="kernel">Kernel configuration</h2> 112<h3 id="kernel-tuning">Tuning kernel/ActivityManager to reduce direct reclaim </h3> 113 114 115 <p>Direct reclaim happens when a process or the kernel tries to allocate a page 116 of memory (either directly or due to faulting in a new page) and the kernel 117 has used all available free memory. This requires the kernel to block the 118 allocation while it frees up a page. This in turn often requires disk I/O to 119 flush out a dirty file-backed page or waiting for <code>lowmemorykiller</code> to kill a 120 process. This can result in extra I/O in any thread, including a UI thread.</p> 121 122 <p>To avoid direct reclaim, the kernel has watermarks that trigger <code>kswapd</code> or 123 background reclaim. This is a thread that tries to free up pages so the next 124 time a real thread allocates it can succeed quickly.</p> 125 126 <p>The default threshold to trigger background reclaim is fairly low, around 2MB 127 on a 2GB device and 636KB on a 512MB device. And the kernel reclaims only a 128 few MB of memory in background reclaim. This means any process that quickly 129 allocates more than a few megabytes is going to quickly hit direct reclaim.</p> 130 131<p>Support for a new kernel tunable is added in the android-3.4 kernel branch as 132 patch 92189d47f66c67e5fd92eafaa287e153197a454f ("add extra free kbytes 133 tunable"). Cherry-picking this patch to a device's kernel will allow 134 ActivityManager to tell the kernel to try to keep 3 full-screen 32 bpp buffers 135 of memory free.</p> 136 137<p>These thresholds can be configured via the framework config.xml</p> 138<p><code> <!-- Device configuration setting the /proc/sys/vm/extra_free_kbytes tunable in the kernel (if it exists). A high value will increase the amount of memory that the kernel tries to keep free, reducing allocation time and causing the lowmemorykiller to kill earlier. A low value allows more memory to be used by processes but may cause more allocations to block waiting on disk I/O or lowmemorykiller. Overrides the default value chosen by ActivityManager based on screen size. 0 prevents keeping any extra memory over what the kernel keeps by default. -1 keeps the default. --><br /> 139<integer name="config_extraFreeKbytesAbsolute">-1</integer></code></p> 140 141<code> 142<p> <!-- Device configuration adjusting the /proc/sys/vm/extra_free_kbytes tunable in the kernel (if it exists). 0 uses the default value chosen by ActivityManager. A positive value will increase the amount of memory that the kernel tries to keep free, reducing allocation time and causing the lowmemorykiller to kill earlier. A negative value allows more memory to be used by processes but may cause more allocations to block waiting on disk I/O or lowmemorykiller. Directly added to the default value chosen by ActivityManager based on screen size. --><br /> 143 <integer name="config_extraFreeKbytesAdjust">0</integer></code> 144 145<h3 id="lowmem">Tuning LowMemoryKiller</h3> 146 147 148 <p>ActivityManager configures the thresholds of the LowMemoryKiller to match its 149 expectation of the working set of file-backed pages (cached pages) required to 150 run the processes in each priority level bucket. If a device has high 151 requirements for the working set, for example if the vendor UI requires more 152memory or if more services have been added, the thresholds can be increased. </p> 153<p>The thresholds can be reduced if too much memory is being reserved for file 154 backed pages, so that background processes are being killed long before disk 155thrashing would occur due to the cache getting too small.</p> 156<p> <code><!-- Device configuration setting the minfree tunable in the lowmemorykiller in the kernel. A high value will cause the lowmemorykiller to fire earlier, keeping more memory in the file cache and preventing I/O thrashing, but allowing fewer processes to stay in memory. A low value will keep more processes in memory but may cause thrashing if set too low. Overrides the default value chosen by ActivityManager based on screen size and total memory for the largest lowmemorykiller bucket, and scaled proportionally to the smaller buckets. -1 keeps the default. --><br /> 157 <integer name="config_lowMemoryKillerMinFreeKbytesAbsolute">-1</integer></code></p> 158<p> <code><!-- Device configuration adjusting the minfree tunable in the lowmemorykiller in the kernel. A high value will cause the lowmemorykiller to fire earlier, keeping more memory in the file cache and preventing I/O thrashing, but allowing fewer processes to stay in memory. A low value will keep more processes in memory but may cause thrashing if set too low. Directly added to the default value chosen by ActivityManager based on screen size and total memory for the largest lowmemorykiller bucket, and scaled proportionally to the smaller buckets. 0 keeps the default. --><br /> 159 <integer name="config_lowMemoryKillerMinFreeKbytesAdjust">0</integer></code></p> 160<h3 id="ksm">KSM (Kernel samepage merging)</h3> 161 162 163 <p>KSM is a kernel thread that runs in the background and compares pages in 164 memory that have been marked <code>MADV_MERGEABLE</code> by user-space. If two pages are 165 found to be the same, the KSM thread merges them back as a single 166 copy-on-write page of memory.</p> 167 168 <p>KSM will save memory over time on a running system, gaining memory duplication 169 at a cost of CPU power, which could have an impact on battery life. You should 170 measure whether the power tradeoff is worth the memory savings you get by 171 enabling KSM.</p> 172 173 <p>To test KSM, we recommend looking at long running devices (several hours) and 174 seeing whether KSM makes any noticeable improvement on launch times and 175 rendering times.</p> 176 177<p>To enable KSM, enable <code>CONFIG_KSM</code> in the kernel and then add the following lines to your` <code>init.<device>.rc</code> file:<br> 178 <code>write /sys/kernel/mm/ksm/pages_to_scan 100<br> 179 write /sys/kernel/mm/ksm/sleep_millisecs 500<br> 180write /sys/kernel/mm/ksm/run 1</code></p> 181<p>Once enabled, there are few utilities that will help in the debugging namely : 182 procrank, librank, & ksminfo. These utilities allow you to see which KSM 183 memory is mapped to what process, which processes use the most KSM memory. 184 Once you have found a chunk of memory that looks worth exploring you can use 185 either the hat utility if it's a duplicate object on the dalvik heap. </p> 186<h3 id="zram">Swap to zRAM</h3> 187 188 189 <p>zRAM swap can increase the amount of memory available in the system by 190 compressing memory pages and putting them in a dynamically allocated swap area 191 of memory.</p> 192 193 <p>Again, since this is trading off CPU time for a small increase in memory, you 194 should be careful about measuring the performance impact zRAM swap has on your 195 system.</p> 196 197 198<p>Android handles swap to zRAM at several levels:</p> 199 200<ul> 201 <li>First, the following kernel options must be enabled to use zRAM swap 202 effectively: 203 <ul> 204 <li><code>CONFIG_SWAP</code></li> 205 <li><code>CONFIG_CGROUP_MEM_RES_CTLR</code></li> 206 <li><code>CONFIG_CGROUP_MEM_RES_CTLR_SWAP</code></li> 207 <li><code>CONFIG_ZRAM</code></li> 208 </ul> 209 </li> 210 <li>Then, you should add a line that looks like this to your fstab:<br /> 211 <code>/dev/block/zram0 none swap defaults zramsize=<size in bytes>,swapprio=<swap partition priority></code><br /> 212 <code><br /> 213 zramsize</code> is mandatory and indicates how much uncompressed memory you want 214 the zram area to hold. Compression ratios in the 30-50% range are usually 215 observed.<br /> 216 <br /> 217 <code>swapprio</code> is optional and not needed if you don't have more than one swap 218 area.<br /> 219 <br /> 220 </li> 221 <li>By default, the Linux kernel swaps in 8 pages of memory at a time. When 222 using ZRAM, the incremental cost of reading 1 page at a time is negligible 223 and may help in case the device is under extreme memory pressure. To read 224 only 1 page at a time, add the following to your init.rc:<br /> 225 `write /proc/sys/vm/page-cluster 0`</li> 226 <li>In your init.rc, after the `mount_all /fstab.X` line, add:<br /> 227 `swapon_all /fstab.X`</li> 228 <li>The memory cgroups are automatically configured at boot time if the 229 feature is enabled in kernel.</li> 230 <li>If memory cgroups are available, the ActivityManager will mark lower 231 priority threads as being more swappable than other threads. If memory is 232 needed, the Android kernel will start migrating memory pages to zRAM swap, 233 giving a higher priority to those memory pages that have been marked by 234 ActivityManager. </li> 235</ul> 236<h3 id="carveouts">Carveouts, Ion and Contiguous Memory Allocation (CMA)</h3> 237 238 <p>It is especially important on low memory devices to be mindful about 239 carveouts, especially those that will not always be fully utilized -- for 240 example a carveout for secure video playback. There are several solutions to 241 minimizing the impact of your carveout regions that depend on the exact 242 requirements of your hardware.</p> 243 <p>If hardware permits discontiguous memory 244 allocations, the ion system heap allows memory allocations from system memory, 245 eliminating the need for a carveout. It also attempts to make large 246 allocations to eliminate TLB pressure on peripherals. If memory regions must 247 be contiguous or confined to a specific address range, the contiguous memory 248 allocator (CMA) can be used.</p> 249<p>This creates a carveout that the system can also 250 use of for movable pages. When the region is needed, movable pages will be 251 migrated out of it, allowing the system to use a large carveout for other 252 purposes when it is free. CMA can be used directly or more simply via ion by 253 using the ion cma heap.</p> 254 255<h2 id="app-opts">Application optimization tips</h2> 256<ul> 257 <li>Review <a 258href="http://developer.android.com/training/articles/memory.html">Managing your 259App's Memory</a> and these past blog posts on the same topic: 260 <ul> 261 <li><a 262href="http://android-developers.blogspot.com/2009/01/avoiding-memory-leaks.html">http://android-developers.blogspot.com/2009/01/avoiding-memory-leaks.html</a></li> 263 <li><a 264href="http://android-developers.blogspot.com/2011/03/memory-analysis-for-android.html">http://android-developers.blogspot.com/2011/03/memory-analysis-for-android.html</a></li> 265 <li><a 266href="http://android-developers.blogspot.com/2009/02/track-memory-allocations.html">http://android-developers.blogspot.com/2009/02/track-memory-allocations.html</a></li> 267 <li> <a 268href="http://tools.android.com/recent/lintperformancechecks">http://tools.android.com/recent/lintperformancechecks</a></li> 269 </ul> 270</li> 271 <li>Check/remove any unused assets from preinstalled apps - 272development/tools/findunused (should help make the app smaller).</li> 273<li>Use PNG format for assets, especially when they have transparent areas</li> 274<li>If writing native code, use calloc() rather than malloc/memset</li> 275<li>Don't enable code that is writing Parcel data to disk and reading it later.</li> 276<li>Don't subscribe to every package installed, instead use ssp filtering. Add 277filtering like below: 278<br /> 279 <code><data android:scheme="package" android:ssp="com.android.pkg1" /><br /> 280 <data android:scheme="package" android:ssp="com.myapp.act1" /></code></li> 281</ul> 282 283<h3 id="process-states">Understand the various process states in Android</h3> 284 285 <ul> 286 <li><p>SERVICE - SERVICE_RESTARTING<br/> 287 Applications that are making themselves run in the background for their own 288 reason. Most common problem apps have when they run in the background too 289 much. %duration * pss is probably a good "badness" metric, although this set 290 is so focused that just doing %duration is probably better to focus on the 291 fact that we just don't want them running at all.</p></li> 292 <li><p>IMPORTANT_FOREGROUND - RECEIVER<br/> 293 Applications running in the background (not directly interacting with the 294 user) for any reason. These all add memory load to the system. In this case 295 the (%duration * pss) badness value is probably the best ordering of such 296 processes, because many of these will be always running for good reason, and 297 their pss size then is very important as part of their memory load.</p></li> 298 <li><p>PERSISTENT<br/> 299 Persistent system processes. Track pss to watch for these processes getting 300 too large.</p></li> 301 <li><p>TOP<br/> 302 Process the user is currently interacting with. Again, pss is the important 303 metric here, showing how much memory load the app is creating while in use.</p></li> 304 <li><p>HOME - CACHED_EMPTY<br/> 305 All of these processes at the bottom are ones that the system is keeping 306 around in case they are needed again; but they can be freely killed at any 307 time and re-created if needed. These are the basis for how we compute the 308 memory state -- normal, moderate, low, critical is based on how many of these 309 processes the system can keep around. Again the key thing for these processes 310 is the pss; these processes should try to get their memory footprint down as 311 much as possible when they are in this state, to allow for the maximum total 312 number of processes to be kept around. Generally a well behaved app will have 313 a pss footprint that is significantly smaller when in this state than when 314 TOP.</p></li> 315 <li> 316 <p>TOP vs. CACHED_ACTIVITY-CACHED_ACTIVITY_CLIENT<em><br/> 317 </em>The difference in pss between when a process is TOP vs. when it is in either 318 of these specific cached states is the best data for seeing how well it is 319 releasing memory when going into the background. Excluding CACHED_EMPTY state 320 makes this data better, since it removes situations when the process has 321 started for some reasons besides doing UI and so will not have to deal with 322 all of the UI overhead it gets when interacting with the user.</p></li> 323 </ul> 324 325 326<h2 id="analysis">Analysis</h2> 327<h3 id="app-startup">Analyzing app startup time</h3> 328 329 330 <p>Use "<code>adb shell am start</code>" with the <code>-P</code> or <code>--start-profiler</code> option to run 331 the profiler when your app starts. This will start the profiler almost 332 immediately after your process is forked from zygote, before any of your code 333is loaded into it.</p> 334<h3 id="bug-reports">Analyze using bugreports </h3> 335 336 337 <p>Now contains various information that can be used for debugging. The services 338 include <code>batterystats</code>, <code>netstats</code>, <code>procstats</code>, and <code>usagestats</code>. You can 339 find them with lines like this:</p> 340 341 342<pre>------ CHECKIN BATTERYSTATS (dumpsys batterystats --checkin) ------ 3437,0,h,-2558644,97,1946288161,3,2,0,340,4183 3447,0,h,-2553041,97,1946288161,3,2,0,340,4183 345</pre> 346<h3 id="persistent">Check for any persistent processes</h3> 347 348 349 <p>Reboot the device and check the processes.<br/> 350 Run for a few hours and check the processes again. There should not be any 351long running processes.</p> 352<h3 id="longevity">Run longevity tests</h3> 353 354 355 <p>Run for longer durations and track the memory of the process. Does it 356 increase? Does it stay constant? Create Canonical use cases and run longevity tests on these scenarios</p> 357