A lot of us heard the word cache and when you ask them about caching they give you a perfect answer but they don’t know how it is built, or on which criteria I should favor this caching framework over that one and so on, in this article we are going to talk about Caching, Caching Algorithms and caching frameworks and which is better than the other.
"Caching is a temp location where I store data in (data that I need it frequently) as the original data is expensive to be fetched, so I can retrieve it faster. "
That what programmer 1 answered in the interview (one month ago he submitted his resume to a company who wanted a java programmer with a strong experience in caching and caching frameworks and extensive data manipulation)
Programmer 1 did make his own cache implementation using hashtable and that what he only knows about caching and his hashtable contains about 150 entry which he consider an extensive data(caching = hashtable, load the lookups in hashtable and everything will be fine nothing else) so lets see how will the interview goes.
: Nice and based on what criteria do you choose your caching solution?
:huh, (thinking for 5 minutes) , mmm based on, on , on the data (coughing…)
: excuse me! Could you repeat what you just said again?
: oh I see, ok list some caching algorithms and tell me which is used for what
: (staring at the interviewer and making strange expressions with his face, expressions that no one knew that a human face can do :D )
: ok, let me ask it in another way, how will a caching behave if it reached its capacity?
: capacity? Mmm (thinking… hashtable is not limited to capacity I can add what I want and it will extend its capacity) (that was in programmer 1 mind he didn’t say it)
The Interviewer thanked programmer 1 (the interview only lasted for 10minutes) after that a woman came and said: oh thanks for you time we will call you back have a nice day
This was the worst interview programmer 1 had (he didn’t read that there was a part in the job description which stated that the candidate should have strong caching background ,in fact he only saw the line talking about excellent package ;) )
After programmer 1 left he wanted to know what were the interviewer talking about and what are the answers to his questions so he started to surf the net, Programmer 1 didn’t know anything else about caching except: when I need cache I will use hashtable
After using his favorite search engine he was able to find a nice caching article and started to read.
Long time ago before caching age user used to request an object and this object was fetched from a storage place and as the object grow bigger and bigger, the user had spend more time to fulfill his request, it really made the storage place suffer a lot coz it had to be working for the whole time this caused both the user and the db to be angry and there were one of 2 possibilities
The user will get upset and complain and even wont use this application again(that was the case always)
The storage place will pack up its bags and leave your application , and that made a big problems(no place to store data) (happened in rare situations )
After few years researchers at IBM (in 60s) introduced a new concept and named it “Cache”
Caching is a temp location where I store data in (data that I need it frequently) as the original data is expensive to be fetched, so I can retrieve it faster.
Caching is made of pool of entries and these entries are a copy of real data which are in storage (database for example) and it is tagged with a
Great so programmer 1 already knows this but what he doesn’t know is caching terminologies which are as follow:
When the client invokes a request (let’s say he want to view product information) and our application gets the request it will need to access the product data in our storage (database), it first checks the cache.
If an entry can be found with a
And the percentage of accesses that result in is known as the or of the cache.
On the contrary when theisn’t found in the cache (no match were found) this is known as , a hit to the back storage is made and the data is fetched back and it is placed in the cache so in future hits it will be found and will make a .
If we encountered a cache miss there can be either a scenarios from two scenarios:
: there is free space in the cache (the cache didn’t reach its limit and there is free space) so in this case the object that cause the will be retrieved from our storage and get inserted in to the cache.
: there is no free space in the cache (cache reached its capacity) so the object that cause will be fetched from the storage and then we will have to decide which object in the cache we need to move in order to place our newly created object (the one we just retrieved) this is done by (caching algorithms) that decide which entry will be remove to make more room which will be discussed below.
When a cache miss occurs, data will be fetch it from the back storage, load it and place it in the cache but how much space the data we just fetched takes in the cache memory? This is known as
When the object that resides in the cache need is updated in the back storage for example it needs to be updated, so keeping the cache up to date is known as
Entry will be invalidate from cache and fetched again from the back storage to get an updated version.
When cache miss happens, the cache ejects some other entry in order to make room for the previously uncached data (in case we don’t have enough room). The heuristic used to select the entry to eject is known as the.
The theoretically optimal page replacement algorithm (also known asor ) is an algorithm that tries to achieve the following: when a cached object need to be placed in the cache, the cache algorithm should replace the entry which will not be used for the longest period of time.
For example, a cache entry that is not going to be used for the next 10 seconds will be replaced by an entry that is going to be used within the next 2 seconds.
Thinking of the optimal replacement policy we can say it is impossible to achieve but some algorithms do near optimal replacement policy based on heuristics.
So everything is based on heuristics so what makes algorithm better than another one? And what do they use for their heuristics?
While reading the article programmer 1 fall a sleep and had nightmare (the scariest nightmare one can ever have)
: nihahha I will invalidate you. (Talking in a mad way)
: no no please let me live, they still need me, I have children.
all cached entries say that before they are invalidated and since when do you have children? Never mind now vanish for ever.
Buhaaahaha , laughed programmer 1 in a scary way, ,silence took over the place for few minutes and then a police serine broke this silence, police caught programmer 1 and he was accused of invalidating an entry that was still needed by a cache client, and he was sent to jail.
Programmer 1 work up and he was really scared, he started to look around and realized that it was just a dream then he continued reading about caching and tried to get rid of his fears.
No one can talk about caching algorithms better than the caching algorithms themselves
I am Least Frequently used; I count how often an entry is needed by incrementing a counter associated with each entry.
I remove the entry with least frequently used counter first am not that fast and I am not that good in adaptive actions (which means that it keeps the entries which is really needed and discard the ones that aren’t needed for the longest period based on the access pattern or in other words the request pattern)
I am Least Recently Used cache algorithm; I remove the least recently used items first. The one that wasn’t used for a longest time.
I require keeping track of what was used when, which is expensive if one wants to make sure that I always discards the least recently used item.
Web browsers use me for caching. New items are placed into the top of the cache. When the cache exceeds its size limit, I will discard items from the bottom. The trick is that whenever an item is accessed, I place at the top.
So items which are frequently accessed tend to stay in the cache. There are two ways to implement me either an array or a linked list (which will have the least recently used entry at the back and the recently used at the front).
I am fast and I am adaptive in other words I can adopt to data access pattern, I have a large family which completes me and they are even better than me (I do feel jealous some times but it is ok) some of my family member are (LRU2 and 2Q) (they were implemented in order to improve LRU caching
I am Least recently used 2, some people calls me least recently used twice which I like it more, I add entries to the cache the second time they are accessed (it requires two times in order to place an entry in the cache); when the cache is full, I remove the entry that has a second most recent access. Because of the need to track the two most recent accesses, access overhead increases with cache size, If I am applied to a big cache size, that would be a problem, which can be a disadvantage. In addition, I have to keep track of some items not yet in the cache (they aren’t requested two times yet).I am better that LRU and I am also adoptive to access patterns.
I am Two Queues; I add entries to an LRU cache as they are accessed. If an entry is accessed again, I move them to second, larger, LRU cache.
I remove entries a so as to keep the first cache at about 1/3 the size of the second. I provide the advantages of LRU2 while keeping cache access overhead constant, rather than having it increase with cache size. Which makes me better than LRU2 and I am also like my family, am adaptive to access patterns.
I am Adaptive Replacement Cache; some people say that I balance between LRU and LFU, to improve combined result, well that’s not 100% true actually I am made from 2 LRU lists, One list, say L1, contains entries that have been seen only once “recently”, while the other list, say L2, contains entries that have been seen at least twice “recently”.
The items that have been seen twice within a short time have a low inter-arrival rate, and, hence, are thought of as “high-frequency”. Hence, we think of L1as capturing “recency” while L2 as capturing “frequency” so most of people think I am a balance between LRU and LFU but that is ok I am not angry form that.
I am considered one of the best performance replacement algorithms, Self tuning algorithm and low overhead replacement cache I also keep history of entries equal to the size of the cache location; this is to remember the entries that were removed and it allows me to see if a removed entry should have stayed and we should have chosen another one to remove.(I really have bad memory)And yes I am fast and adaptive.
I am most recently used, in contrast to LRU; I remove the most recently used items first. You will ask me why for sure, well let me tell you something when access is unpredictable, and determining the least most recently used entry in the cache system is a high time complexity operation, I am the best choice that’s why.
I am so common in the database memory caches, whenever a cached record is used; I replace it to the top of stack. And when there is no room the entry on the top of the stack, guess what? I will replace the top most entry with the new entry.
I am first in first out; I am a low-overhead algorithm I require little effort for managing the cache entries. The idea is that I keep track of all the cache entries in a queue, with the most recent entry at the back, and the earliest entry in the front. When there e is no place and an entry needs to be replaced, I will remove the entry at the front of the queue (the oldest entry) and replaced with the current fetched entry. I am fast but I am not adaptive
Hello I am second change I am a modified form of the FIFO replacement algorithm, known as the Second chance replacement algorithm, I am better than FIFO at little cost for the improvement. I work by looking at the front of the queue as FIFO does, but instead of immediately replacing the cache entry (the oldest one), i check to see if its referenced bit is set(I use a bit that is used to tell me if this entry is being used or requested before or no). If it is not set, I will replace this entry. Otherwise, I will clear the referenced bit, and then insert this entry at the back of the queue (as if it were a new entry) I keep repeating this process. You can think of this as a circular queue. Second time I encounter the same entry I cleared its bit before, I will replace it as it now has its referenced bit cleared. am better than FIFO in speed
am Clock and I am a more efficient version of FIFO than Second chance because I don’t push the cached entries to the back of the list like Second change do, but I perform the same general function as Second-Chance.
I keep a circular list of the cached entries in memory, with the "hand" (something like iterator) pointing to the oldest entry in the list. When cache miss occurs and no empty place exists, then I consult the R (referenced) bit at the hand's location to know what I should do. If R is 0, then I will place the new entry at the "hand" position, otherwise I will clear the R bit. Then, I will increment the hand (iterator) and repeat the process until an entry is replaced.I am faster even than second chance.
I am simple time-based caching; I invalidate entries in the cache based on absolute time periods. I add Items to the cache, and they remain in the cache for a specific amount of time. I am fast but not adaptive for access patterns
I am extended time based expiration cache, I invalidate the items in the cache is based on relative time periods. I add Items the cache and they remain in the cache until I invalidate them at certain points in time, such as every five minutes, each day at 12.00.
I am Sliding time-base expiration, I invalidate entries a in the cache by specifying the amount of time the item is allowed to be idle in the cache after last access time. after that time I will invalidate it . I am fast but not adaptive for access patterns
Ok after we listened to some replacement algorithms (famous ones) talking about themselves, some other replacement algorithms take into consideration some other criteria like:
: if items have different costs, keep those items that are expensive to obtain, e.g. those that take a long time to get.
: If items have different sizes, the cache may want to discard a large item to store several smaller ones.
: Some caches keep information that expires (e.g. a news cache, a DNS cache, or a web browser cache). The computer may discard items because they are expired. Depending on the size of the cache no further caching algorithm to discard items may be necessary.
After programmer 1 did read the article he thought for a while and decided to send a mail to the author of this caching article, he felt like he heard the author name before but he couldn’t remember who this person was but anyway he sent him mail asking him about what if he has a distributed environment? How will the cache behave?
The author of the caching article got his mail and ironically it was the man who interviewed programmer 1 :D, The author replied and said :
*Caching Data can be stored in separate memory area from the caching directory itself (who handle the caching entries and so on) can be across network or disk for example.
*Distrusting the cache allows increase in the cache size.
*In this case the retrieval cost will increase also due to network request time.
*This will also lead to hit ratio increase due to the large size of the cache
But how will this work?
Let’s assume that we have 3 Servers, 2 of them will handle the distributed caching (have the caching entries), and the 3rd one will handle all the requests that are coming (Which asks about cached entries):
Step 1: the application requests keys entry1, entry2 and entry3, after resolving the hash values for these entries, and based on the hashing value it will be decided to forward the request to the proper server.
Step 2: the main node sends parallel requests to all relevant servers (which has the cache entry we are looking for).
Step 3: the servers send responses to the main node (which sent the request in the 1st place asking to the cached entry).
Step 4: the main node sends the responses to the application (cache client).
*And in case the cache entry were not found (the hashing value for the entry will be still computed and will redirect either to server 1 or server 2 for example and in this case our entry wont be found in server 1 so it will fetched from the DB and added to server 1 caching list.
Most caches can be evaluated based on measuring the hit ratio and comparing to the theoretical optimum, this is usually done by generation a list of cache keys with no real data, but here the hit ratio measurement assumes that all entries have the same retrieval cost which is not true for example in web caching the number of bytes the cache can server is more important than the number of hit ration (for example I can replace the big entry will 10 small entries which is more effective in web)
We have seen some of popular algorithms that are used in caching, some of them are based on time, cache object size and some are based on frequency of usage, next part we are going to talk about the caching framework and how do they make use of these caching algorithms, so stay tuned ;)