You don’t have to be a software engineer for long until you hit into performance issues. The database loads too slow. Calculating the route takes forever. Remote calls hang. What to do?
SQL query optimization? JS packing? Indices? Algorithm optimizations? Multi threading? Change to C++? Cloud? Micro services? Sharding? NoSQL? Remove features? Increase hardware requirements? Kernel optimizations? JIT loading? Load balancing? Remove variables? Use Protobuf? Better Compression? Change libraries? Change frameworks? GC tuning?
There are a lot of ways speed up code and not all of them will be effective or worth the effort. But what’s the best option in your case? This is a vast specialized area where for example all the techniques listed above have their own experts if you happen to need one. Before you go and hire one, let’s go through some steps that may help you decide.
Step 1: Diagnosis
Diagnosing the problem is the obvious first step. It can be as simple as timing function calls, but usually it’s a better to fire up a profiler, something like the Chrome Developer Tools Profiles Pane or Java Mission Control, whatever the platform you’re on. There are a lot of performance analysis tools out there so go and find the one that is applicable to your needs. Take the time to learn how to use one and they’ll usually pinpoint you to the worst offender rather quickly.
This is also a good time to spot programming errors like memory leaks. Those need to be fixed regardless. Look for object counts and allocated memory and see if the numbers make sense.
It’s almost a certainty that the Pareto principle will hold — 80% of the slowness is caused by 20% of the code. In fact it’s usually much more extreme than that — a single often hit function could be the bottle neck for the entire 100k line code base so most of your effort should go to fix that first.
Often the problem is as obvious as the fix, so grabbing these low hanging fruit can solve most of your woes. If it’s not obvious how and what to fix, keep reading.
Step 2: Understand Your Domain
Even if something could be done it doesn’t mean it should. Knowing your domain, the given hardware limitations and the expected load is vital before you decide to do anything drastic. Are we talking about a single player game here? A hardware sensor reading app that runs on your Raspberry Pi? A business app with max 100 users? A web app with millions of concurrent users? Changing from Ruby to C might make sense on the Pi, but probably not with your Rails app.
Let’s take the the Finnish Parliament and it’s voting system problem for example:
The Finnish parliament upgraded its electronic voting system in the early 2000s. In the final phases of testing they found out that when enough members pressed their voting buttons at the same time, the system couldn’t handle the load and crashed.
In the Finnish Parliament the chair calls for a vote on something, and you have 200 members of parliament voting for something by pressing one three options, Jaa (Yes), Ei (No) or Tyhjää (None of the above). They usually have something like 10 seconds or so to do this. So if you think about it, the scenario that didn’t work is the most common use case — most of the members will press their button immediately as the voting is up.
So what can we learn from this, at least performance wise? Understand your domain. Changing to a serverless cloud architecture probably won’t help if the problem is with the electronics, even if the back end was kinda slow. Aim your goals to match the expected load now and in the near future. There’s no need to make the voting system work with 10000 concurrent users. The parliament won’t (I hope!) grow to that size. In fact to set the goal to exactly 200 is fine, maybe throw an extra 20 to have a margin of error.
Loads aren’t usually this predictable and static. There’s increases and peaks and that surprise hoard of Reddit front page users. But still, know your playing field and accept that it might not be worth it to prepare for everything at once. You’ll never ship if you set your Android app performance goal to the slowest phone out there. Trust me, it’s slower than that sloth.
If you have a hundred users and hoping to reach a thousand the measures you should take now are different compared to getting from thousand to a million. The latter might require techniques like using micro services and they will add a huge layer of complexity to the system with added risk. Getting to a thousand is likely a matter of optimizing your database queries.
Step 3: Decide Course of Action
So the app is slow and you’ve diagnosed the problem. It’s that function that loads the data from the remote database. So how should you proceed?
Here’s a domain agnostic questionnaire to narrow your options:
Can I skip doing this completely?
Seriously think about it. Even your ultra optimized algorithm is slower than my deleted code. Many of the greatest performance tricks like adaptive tile refresh are at their heart ways to avoid doing stuff. Think about the use case and if it’s not vital, maybe you can do without it,
Maybe some light weight version of it will suffice? Do you really really need the 7MB js library to load the 10MB SVG company logo that’s 10 pixels across in the corner?
Can I move it somewhere less expensive?
Maybe you have a computationally expensive list on the front page. Does everyone use it? If not, perhaps move it somewhere so only the people that really need it are inflicted. It can also save cycles on the back end.
Can I do it only once?
Also known as caching. Doing something only once and then using the result is effective, but it’s also hard. Especially if you invent your own caching scheme, you’ll open a huge can of worms to deal with. Your existing test suite might also be inadequate to handle it when calls suddenly return outdated data. Be warned of regressions. Tread carefully and prefer any existing technologies like ehcache out there.
If your application has a long chain of computations, also consider caching the intermediary results. If you have the memory, use it.
Can I do it partially?
This can be as simple as paging, or it can be limiting the amount of data shown to the user and requesting more data if the use requests it. See also JIT.
Can I index it?
Indices are simplified an added data structure that speed up data retrieval. If your database queries are slow, SQL query optimization usually involves adding indices. But it can be an added dictionary in your code that stores keys that speed up results. For example if you have a graph of cities on a map, there can be an additional dictionary(the index) of city names that can be used to retrieve the nodes.
The cost is maintaining the index and this usually means increased storage requirements and slower writes.
Can I JIT?
JIT (just in time) means delaying the loading of the data until it’s really required. This can be some form of lazy loading in the code (the object is initialized the moment it’s actually used) or having a scrolling list that loads more elements when the user scrolls down to the, such as the RecyclerView. Almost all modern languages and frameworks have a variant of lazy loading, so look into it.
Can I preload?
The opposite of JIT is loading a chunk data in advance. Is it more tolerable to have a single long pause as opposed to smaller stutters here and there? If for example the hardware is limited, constant loading can be out of the question and a single long wait might be worth it if the UI is unusable garbage otherwise.
Can I do it secretly?
Splash screens. Everyone loves splash screens right? Well that’s not so secret, but doing stuff on the background when the user does something else can put the problem under the rug. Things like BackgroundWorkers are good for this.
The downside is synchronization —does someone else need the result of the computation? This will inevitably add complexity and introduce fun risks like dead locks.
Can I do it concurrently?
This beast is also called multithreading. Can I split the problem into parts so that you can solve them in parallel and then gather the results? This can be very effective, but it’s also hard. Very hard.
Deadlocks, race conditions, Heisenbugs, synchronization issues, thread safety, testing challenges.. Fun times. If you’re not familiar with this already, turn back.
Can I increase memory/cpu?
Sometimes there’s nothing wrong per se with the app, it just runs out of resources. If the OS runs out memory and it swaps, it’s going to be detrimental. Same goes for CPU cycles. If the app doesn’t leak memory, adding memory is probably the cheapest choice to fix your problems. There’s a lot to be done optimizing your OS if you know the problem lies there.
In case of web apps for example, you can also split the load onto multiple computers and use a load balancer. In the age of cloud computing, this can also can scale like magic and you do basically nothing (but pay).
Can I guess ahead?
Buffering.. Video is the obvious example, or Spectre err.. I mean branch prediction in CPU architectures. A combination of doing things secretly and trying to guess what the user is going to do next.
If you guess wrong, you waste cycles (or create security vulnerabilities).
Can I tolerate errors?
With really big data sets, the eventual consistency of NoSQL databases is also an example of this. Being optimistic and fixing the errors later can be a significant performance boost.
Can I optimize the algorithm?
There’s never a shortage of things to do on the algorithm level. For example, look at the known subclasses of the Collection interface in Java. Do you remember all of them and know when to use them? Or is everything an ArrayList in your app? On top of that there’s stuff like fastutil if you really need to push it. It’s not uncommon to see code needlessly doing the same thing over and over and looping through a long list when a HashMap would’ve worked.
There’s a lot to be done on the code level, and if your unit tests are solid, it should also be relatively safe. Again, YMMV.
Can I reduce requests?
Fetching data from a remote server is slow, very slow compare to your computer memory. The round trip to your data center more than 100 000x slower than to your RAM. If you make a lot of round trips, can you make a single larger batch request for example?
The downside of batching can be that the round trip time of a single request can increase, if the batching module waits for more data or the request size factors into it. YMMV.
Can I use a dedicated text searching service?
Does your application have an internal text search that’s slow? Things like Lucene can feel like black magic if you have the memory to spare. Use them if you have the memory to spare.
Can I split the database into more manageable chunks?
Also known as sharding. If your database has grown to a gigantic size perhaps this is an option. It comes with the caveat of added complexity.
Can I compress it?
Can I use something like webpack to minimize the request load in my web app? Or use zip compression in large network transfers? You can also look into things like Protocol Buffers if you are sure that the REST JSON overhead is too much.
Can I change the framework or programming language?
I don’t know, can you? Would it help? Unless you’re on the Pi or similar limited hardware, this should be at the bottom of your list.
Can I run serverless?
It solves all these problems and all the cool kids do this, so we should too, right? Well, maybe. I have no experience of this, tell me how it went if you do this!
Sometimes the problem is much more hairy — it persists only under heavier loads harder to reproduce or only on that gremlin infected machine that looks identical to others. Or you’ve simply run out of the obvious stuff to optimize. This type of stuff can get very specialized, you’ll save time by hiring an expert to figure out out. Doing things like flame graphs can help visualize the painful areas.
Whatever you decide to do, make sure you have tests in place that you can verify that the changes work and don’t cause regressions. Many of these techniques can really change the internals quite dramatically, even if nothing changes on the outside. Change incrementally, one optimization at a time if possible.