Caching stresses certain hardware subsystems more than others. Although the key to good cache performance is good overall system performance, the following list is arranged in order of decreasing importance:
Disk random seek time
Amount of system memory
Sustained disk throughput
Do not drastically underpower any one subsystem, or performance will suffer. In the case of catastrophic hardware failure you must have a ready supply of alternate parts. When your cache is critical, you should have a (working!) standby machine with operating system and Squid installed. This can be kept ready for nearly instantaneous swap-out. This will, of course, increase your costs, something that you may want to take into account. Chapter 13 covers standby procedures in detail.
When deciding on your cache's horsepower, many factors must be taken into account. To decide on your machine, you need an idea of the load that it will need to sustain: the peak number of requests per minute. This number indicates the number of 'objects' downloaded in a minute by clients, and can be used to get an idea of your cache load.
Computing the peak number of requests is difficult, since it depends on the browsing habits of users. This, in turn, makes deciding on the required hardware difficult. If you don't have many statistics as to your Internet usage, it is probably worth your while installing a test cache server (on any machine that you have handy) and pointing some of your staff at it. Using ratios you can estimate the number of requests with a larger user base.
When gathering statistics, make sure that you judge the 'peak' number of requests, rather than an average value. You shouldn't take the number of requests per day and divide, since your peak (during, for example, lunch hour) can be many times your average number of requests.
It's a very good idea to over-estimate hardware requirements. Stay ahead of the growth curve too, since an overloaded cache can spiral out of control due to a transient network problems If a cache cannot deal with incoming requests for some reason (say a DNS outage), it still continues to accept incoming requests, in the hope that it can deal with them. If no requests can be handled, the number of concurrent connections will increase at the rate that new requests arrive.
If your cache runs close to capacity, a temporary glitch can increase the number of concurrent, waiting, requests tremendously. If your cache can't cope with this number of established connections, it may never be able to recover, with current connections never being cleared while it tries to deal with a huge backlog.
Squid 2.0 may be configured to use threads to perform asynchronous Input/Output on operating systems that supports Posix threads. Including async-IO can dramatically reduce your cache latency, allowing you to use a less powerful machine. Unfortunately not all systems support Posix threads correctly, so your choice of hardware can depend on the abilities of your operating system. Your choice of operating system is discussed in the next section - see if your system will support threads there.
There are numerous things to consider when buying disks. Earlier on we mentioned the importance of disks with a fast random-seek time, and with high sustained-throughput. Having the world's fastest drive is not useful, though, if it holds a tiny amount of data. To cache effectively you need disks that can hold a significant amount of downloaded data, but that are fast enough to not slow your cache to a crawl.
Seek time is one of the most important considerations if your cache is going to be loaded. If you have a look at a disk's documentation there is normally a random seek time figure. The smaller this value the better: it is the average time that the disk's heads take to move from a random track to another (in milliseconds). Operating systems do all sorts of interesting things (which are not covered here) to attempt to speed up disk access times: waiting for disks can slow a machine down dramatically. These operating system features make it difficult to estimate how many requests per second your cache can handle before being slowed by disk access times (rather than by network speed). In the next few paragraphs we ignore operating system readahead, inode update seeks and more: it's a back of the envelope approximation for your use.
If your cache does not use asynchronous Input-Output (described in the Operating system section shortly) then your cache loses a lot of the advantage gained by multiple disks. If your cache is going to be loaded (or is running anywhere approaching capacity according to the formulae below) you must ensure that your operating system supports posix threads!
A cache with one disk has to seek at least once per request (ignoring
RAM caching of the disk and inode update times). If you have only one disk,
the formula for working out seeks per second (and hence requests per
second) is quite simple:
requests per second = 1000/seek time 1000
Squid load-balances writes between multiple cache disks, so if you
have more than one data disk your seeks-per-second per disk will be lower.
Almost all operating systems will increase random seek time in a
semi-linear fashion as you add more disks, though others may have a small
performance penalty. If you add more disks to the equation, the requests per second value becomes even more approximate! To simplify things in the
meantime, we are going to assume that you use only disks with the same seek
time. Our formula thus becomes:
theoretical requests per second = -----------------
(seek time)/(number of disks)
requests per second = 1000/seek time
Let's consider a less theoretical example:
I have three disks - all have 12ms seek times. I can thus (theoretically,
as always) handle:
requests per second = 1000/(12/3) = 1000/4 = 250 requests per second
requests per second = 1000/(12/3) = 1000/4 = 250 requests per second
While we are on this topic: many people query the use of IDE disks in caches. IDE disks these days generally have very similar seek times to SCSI disks, and (with DMA-compatible IDE controllers) approach the speed of data transfer without slowing the whole machine down.
Deciding how much disk space to allocate to Squid is difficult. For the pilot project you can simply allocate a few megabytes, but this is unlikely to be useful on a production cache.
The amount of disk space required depends on quite a few factors.
Assume that you were to run a cache just for yourself. If you were to allocate 1 gig of disk, and you browse pages at a rate of 10 megabytes per day, it will take at least 100 days for you to fill the cache.
You can thus see that the rate of incoming cache queries influences the amount of disk to allocate.
If you examine the other end of the scale, where you have 10 megabytes of disk, and 10 incoming queries per second, you will realize that at this rate your disk space will not last very long. Objects are likely to be pushed out of the cache as they arrive, so getting a hit would require two people to be downloading the object at almost exactly the same time. Note that the latter is definitely not impossible, but it happens only occasionally on loaded caches.
The above certainly appears simple, but many people do not extrapolate. The same relationships govern the expulsion of objects from your cache at larger cache store sizes. When deciding on the amount of disk space to allocate, you should determine approximately how much data will pass through the cache each day. If you are unable to determine this, you could simply use your theoretical maximum transfer rate of your line as a basis. A 1mb/s line can transfer about 125000 bytes per second. If all clients were setup to access the cache, disk would be used at about 125k per second, which translates to about 450 megabytes per hour. If the bulk of your traffic is transferred during the day, you are probably transferring 3.6 gigabytes per day. If your line was 100% used, however, you would probably have upgraded it a while ago, so let's assume you transfer 2 gigabytes per day. If you wanted to keep ALL data for a day, you would have to have 2 gigabytes of disk for Squid.
The feasibility of caching depends on two or more users visiting the same page while the object is still on disk. This is quite likely to happen with the large sites (search engines, and the default home pages in respective browsers), but the chances of a user visiting the same obscure page is slim, simply due to the volume of pages. In many cases the obscure pages are on the slowest links, frustrating users. Depending on the number of users requesting pages you should keep pages for longer, so that the chances of different users accessing the same page twice is higher. Determining this value, however, is difficult, since it also depends on the average object size, which, in turn, depends on user habits.
Some people use RAID systems on their caches. This can dramatically increase availability, but a RAID-5 system can reduce disk throughput significantly. If you are really concerned with uptime, you may find a RAID system useful. Since the actual data in the cache store is not vital, though, you may prefer to manually fail-over the cache, simply re-formatting or replacing drives. Sure, your cache may have a lower hit-ratio for a short while, but you can easily balance this minute cost against what hardware to do automatic failover would have cost you.
You should probably base your purchase on the bandwidth description above, and use the data discussed in chapter 11 to decide when to add more disk.
Squid keeps an in-memory table of objects in RAM. Because of the way that Squid checks if objects are in the file store, fast access to the table is very important. Squid slows down dramatically when parts of the table are in swap.
Since Squid is one large process, swapping is particularly bad. If the operating system has to swap data, Squid is placed on the 'sleeping tasks' queue, and cannot service other established connections. (? hmm. it will actually get woken up straight away. I wonder if this is relevant ?)
Each object stored on disk uses about 75 bytes (? get exact value ?) of RAM in the index. The average size of an object on the Internet is about 13kb, so if you have a gigabyte of disk space you will probably store around about 80 000 objects.
At 75 bytes of RAM per object, 80 000 objects require about six megabytes of RAM. If you have 8gigs of disk you will need 48Mb of RAM just for the object index. It is important to note that this excludes memory for your operating system, the Squid binary, memory for in-transit objects and spare RAM for for disk cache.
So, what should your sustained-thoughput of your disks be? Squid tends to read in small blocks, so throughput is of lesser importance than random seek times. Generally disks with fast seeks are high throughput, and most disks (even IDE disks these days) can transfer data faster than clients can download it from you. Don't blow a year's budget on really high-speed disks, go for lower-seek times instead - or add more disks.
Squid is not generally CPU intensive. On startup Squid can use a lot of CPU while it works out what is in the cache, and a slow CPU can slow down access to the cache for the first few minutes upon startup. A Pentium 133 machine generally runs pretty idle, while receiving 7 TCP requests a second A multiprocessor machine generally doesn't increase speed dramatically: only certain portions of the Squid code are threaded. These sections of code are not processor intensive either: they are the code paths where Squid is waiting for the operating system to complete something. A multiprocessor machine generally does not reduce these wait times: more memory (for caching of data) and more disks may help more.