Cache Misses L1 < L2 < L3 - caching

I have a specific piece of software that exhibits a behavior were the miss ratios look like this:
L1-dcache-misses < L2-misses< L3-misses
How can this be the case?
The miss ratios are computed using perf by looking at the refill counters divided by the total number of accesses for each cache in part.

L1-dcache-misses is the fraction of all loads that miss in L1d cache.
L2-misses is the fraction of requests that make it to L2 at all (miss in L1) and then miss in L2. Similar for L3.
An L1d hit isn't part of the total L2 accesses. (Which makes sense because L2 never even sees it).
This is pretty normal for a workload with good locality over a small working set, but the accesses that miss in L1d have poor spatio-temporal locality and tend to miss in outer caches as well.
L1d filters out all the "easy" very-high-locality accesses, leaving L2 and L3 to only deal with the "harder" accesses. You can say that L1d exists to give excellent latency (and bandwidth) for the smallest hottest working set, while L2 tries to catch stuff that falls through the cracks. Then L3 only sees the "most difficult" parts of your access pattern.
Also, if you're on an Intel CPU, note that perf doesn't just use mem_load_retired.l1_miss events and so on; it tries to count multiple misses to the same line of L1d as a single miss by using the L1D.REPLACEMENT event. LLC-loads and load-misses use OFFCORE_RESPONSE events, not mem_load_retired.l3_hit / miss. See
How does Linux perf calculate the cache-references and cache-misses events
(Two loads to the same cache line that isn't ready yet will share the same LFB to track the incoming line, so this accounting makes sense. Also if we care about lines touched / missed instead of individual loads. But L1-dcache-loads uses MEM_INST_RETIRED.ALL_LOADS which does count every load. So not even the perf-reported L1 hit rate is not really the per-instruction L1d load hit rate. It will be higher for any program with spatial locality in its L1d misses.)

Related

why L2 cache utilization is low even though L2 cache hit rate is about 93% in cuda?

I'm doing expriments by using cuda.
I thought that if L2 cache hit ratio is high, performance will increase.
However, from nvprof, L2 cache utilization is low even though L2 cache hit rate is about 93%.
Why this happens? Are there examples that make it happen?
They are different things. You can look for data in L2 cache infrequently, yet when you do so find that L2 almost always contains the data you are looking for.
In fact, this is a combination that is quite likely to go together, because low utilization means useful data is less likely to get evicted from the cache.
L2 utilization and hit rate are orthogonal concepts.
L2 utilization % measures how many operations (reads/writes/atomics) the L2 cache performed, compared to its peak performance. You can alternatively think of this as a proxy for "how much L2 bandwidth did I use" given there is a fixed bandwidth between L1 and L2 on a given GPU. Note, this metric is not measuring the % of L2 capacity used. (to simplify, in the diagram below, think of it as measuring the throughput of arrows next to the red dots)
L2 cache hit rate measures when an L1 miss occurs, how often was it found in L2. (in the diagram, think of L2 cache tags at the green X)
Original diagram from Dissecting the NVidia Turing T4 GPU via Microbenchmarking
Hypothetically:
Some CUDA kernel could read a single L1 cacheline (128B) per SM once, incurring a single L2 read that always hits. The L2 utilization would be ~0%, with L2 hit-rate of 100%.
A different CUDA kernel could achieve ~100% L2 utilization and 100% L2 hit-rate, by performing tons of loads that either miss in L1, or were "cache global" loads, where the set of accessed addresses fit within the size of the L2.
Yet another CUDA kernel could achieve high L2 utilization and low L2 hit-rate, by performing tons of loads that either miss in L1, or were "cache global" loads that are scattered throughout a Gigabyte sized buffer (i.e. that don't all fit simultaneously in L2).
See also
The tables of metrics in the CUDA Toolkit Profiler documentation.
Dissecting the NVidia Turing T4 GPU via Microbenchmarking
Dissecting the NVIDIA Volta GPU Architecture via Microbenchmarking

Optimal buffer size to avoid cache misses for recent i7 / i9 CPUs

Let's assume an algorithm is repeatedly processing buffers of data, it may be accessing say 2 to 16 of these buffers, all having the same size. What would you expect to be the optimum size of these buffers, assuming the algorithm can process the full data in smaller blocks.
I expect the potential bottleneck of cache misses if the blocks are too big, but of course the bigger the blocks the better for vectorization.
Let's expect current i7/i9 CPUs (2018)
Any ideas?
Do you have multiple threads? Can you arrange things so the same thread uses the same buffer repeatedly? (i.e. keep buffers associated with threads when possible).
Modern Intel CPUs have 32k L1d, 256k L2 private per-core. (Or Skylake-AVX512 has 1MiB private L2 caches, with less shared L3). (Which cache mapping technique is used in intel core i7 processor?)
Aiming for L2 hits most of the time is good. L2 miss / L3 hit some of the time isn't always terrible, but off-core is significantly slower. Remember that L2 is a unified cache, so it covers code as well, and of course there's stack memory and random other demands for L2. So aiming for a total buffer size of around half L2 size usually gives a good hit-rate for cache-blocking.
Depending on how much bandwidth your algorithm can use, you might even aim for mostly L1d hits, but small buffers can mean more startup / cleanup overhead and spending more time outside of the main loop.
Also remember that with Hyperthreading, each logical core competes for cache on the physical core it's running on. So if two threads end up on the same physical core, but are touching totally different memory, your effective cache sizes are about half.
Probably you should make the buffer size a tunable parameter, and profile with a few different sizes.
Use perf counters to check if you're actually avoiding L1d or L2 misses or not, with different sizes, to help you understand whether your code is sensitive to different amounts of memory latency or not.

Why do L1 and L2 Cache waste space saving the same data?

I don't know why L1 Cache and L2 Cache save the same data.
For example, let's say we want to access Memory[x] for the first time. Memory[x] is mapped to the L2 Cache first, then the same data piece is mapped to L1 Cache where CPU register can retrieve data from.
But we have duplicated data stored on both L1 and L2 cache, isn't it a problem or at least a waste of storage space?
I edited your question to ask about why CPUs waste cache space storing the same data in multiple levels of cache, because I think that's what you're asking.
Not all caches are like that. The Cache Inclusion Policy for an outer cache can be Inclusive, Exclusive, or Not-Inclusive / Not-Exclusive.
NINE is the "normal" case, not maintaining either special property, but L2 does tend to have copies of most lines in L1 for the reason you describe in the question. If L2 is less associative than L1 (like in Skylake-client) and the access pattern creates a lot of conflict misses in L2 (unlikely), you could get a decent amount of data that's only in L1. And maybe in other ways, e.g. via hardware prefetch, or from L2 evictions of data due to code-fetch, because real CPUs use split L1i / L1d caches.
For the outer caches to be useful, you need some way for data to enter them so you can get an L2 hit sometime after the line was evicted from the smaller L1. Having inner caches like L1d fetch through outer caches gives you that for free, and has some advantages. You can put hardware prefetch logic in an outer or middle level of cache, which doesn't have to be as high-performance as L1. (e.g. Intel CPUs have most of their prefetch logic in the private per-core L2, but also some prefetch logic in L1d).
The other main option is for the outer cache to be a victim cache, i.e. lines enter it only when they're evicted from L1. So you can loop over an array of L1 + L2 size and probably still get L2 hits. The extra logic to implement this is useful if you want a relatively large L1 compared to L2, so the total size is more than a little larger than L2 alone.
With an exclusive L2, an L1 miss / L2 hit can just exchange lines between L1d and L2 if L1d needs to evict something from that set.
Some CPUs do in fact use an L2 that's exclusive of L1d (e.g. AMD K10 / Barcelona). Both of those caches are private per-core caches, not shared, so it's like the simple L1 / L2 situation for a single core CPU you're talking about.
Things get more complicated with multi-core CPUs and shared caches!
Barcelona's shared L3 cache is also mostly exclusive of the inner caches, but not strictly. David Kanter explains:
First, it is mostly exclusive, but not entirely so. When a line is sent from the L3 cache to an L1D cache, if the cache line is shared, or is likely to be shared, then it will remain in the L3 – leading to duplication which would never happen in a totally exclusive hierarchy. A fetched cache line is likely to be shared if it contains code, or if the data has been previously shared (sharing history is tracked). Second, the eviction policy for the L3 has been changed. In the K8, when a cache line is brought in from memory, a pseudo-least recently used algorithm would evict the oldest line in the cache. However, in Barcelona’s L3, the replacement algorithm has been changed to also take into account sharing, and it prefers evicting unshared lines.
AMD's successor to K10/Barcelona is Bulldozer. https://www.realworldtech.com/bulldozer/3/ points out that Bulldozer's shared L3 is also victim cache, and thus mostly exclusive of L2. It's probably like Barcelona's L3.
But Bulldozer's L1d is a small write-through cache with an even smaller (4k) write-combining buffer, so it's mostly inclusive of L2. Bulldozer's write-through L1d is generally considered a mistake in the CPU design world, and Ryzen went back to a normal 32kiB write-back L1d like Intel has been using all along (with great results). A pair of weak integer cores form a "cluster" that shares an FPU/SIMD unit, and shares a big L2 that's "mostly inclusive". (i.e. probably a standard NINE). This cluster thing is Bulldozer's alternative to SMT / Hyperthreading, which AMD also ditched for Ryzen in favour of normal SMT with a massively wide out-of-order core.
Ryzen also has some exclusivity between core clusters (CCX), apparently, but I haven't looked into the details.
I've been talking about AMD first because they have used exclusive caches in recent designs, and seem to have a preference for victim caches. Intel hasn't tried as many different things, because they hit on a good design with Nehalem and stuck with it until Skylake-AVX512.
Intel Nehalem and later use a large shared tag-inclusive L3 cache. For lines that are modified / exclusive (MESI) in a private per-core L1d or L2 (NINE) cache, the L3 tags still indicate which cores (might) have a copy of a line, so requests from one core for exclusive access to a line don't have to be broadcast to all cores, only to cores that might still have it cached. (i.e. it's a snoop filter for coherency traffic, which lets CPUs scale up to dozens of cores per chip without flooding each other with requests when they're not even sharing memory.)
i.e. L3 tags hold info about where a line is (or might be) cached in an L2 or L1 somewhere, so it knows where to send invalidation messages instead of broadcasting messages from every core to all other cores.
With Skylake-X (Skylake-server / SKX / SKL-SP), Intel dropped that and made L3 NINE and only a bit bigger than the total per-core L2 size. But there's still a snoop filter, it just doesn't have data. I don't know what Intel's planning to do for future (dual?)/quad/hex-core laptop / desktop chips (e.g. Cannonlake / Icelake). That's small enough that their classic ring bus would still be great, so they could keep doing that in mobile/desktop parts and only use a mesh in high-end / server parts, like they are in Skylake.
Realworldtech forum discussions of inclusive vs. exclusive vs. non-inclusive:
CPU architecture experts spend time discussing what makes for a good design on that forum. While searching for stuff about exclusive caches, I found this thread, where some disadvantages of strictly inclusive last-level caches are presented. e.g. they force private per-core L2 caches to be small (otherwise you waste too much space with duplication between L3 and L2).
Also, L2 caches filter requests to L3, so when its LRU algorithm needs to drop a line, the one it's seen least-recently can easily be one that stays permanently hot in L2 / L1 of a core. But when an inclusive L3 decides to drop a line, it has to evict it from all inner caches that have it, too!
David Kanter replied with an interesting list of advantages for inclusive outer caches. I think he's comparing to exclusive caches, rather than to NINE. e.g. his point about data sharing being easier only applies vs. exclusive caches, where I think he's suggesting that a strictly exclusive cache hierarchy might cause evictions when multiple cores want the same line even in a shared/read-only manner.

When L1 misses are a lot different than L2 accesses... TLB related?

I have been running some benchmarks on some algorithms and profiling their memory usage and efficiency (L1/L2/TLB accesses and misses), and some of the results are quite intriguing for me.
Considering an inclusive cache hierarchy (L1 and L2 caches), shouldn't the number of L1 cache misses coincide with the number of L2 cache accesses? One of the explanations I find would be TLB related: when a virtual address is not mapped in TLB, the system automatically skips searches in some cache levels.
Does this seem legitimate?
First, inclusive cache hierarchies may not be so common as you assume. For example, I do not think any current Intel processors - not Nehalem, not Sandybridge, possibly Atoms - have an L1 that is included within the L2. (Nehalem and probably Sandybridge do, however, have both L1 and L2 included within L3; using Intel's current terminology, FLC and MLC in LLC.)
But, this doesn't necessarily matter. In most cache hierarchies if you have an L1 cache miss, then that miss will probably be looked up in the L2. Doesn't matter if it is inclusive or not. To do otherwise, you would have to have something that told you that the data you care about is (probably) not in the L2, you don't need to look. Although I have designed protocols and memory types that do this - e.g. a memory type that cached only in the L1 but not the L2, useful for stuff like graphics where you get the benefits of combining in the L1, but where you are repeatedly scanning over a large array, so caching in the L2 not a good idea. Bit I am not aware of anyone shipping them at the moment.
Anyway, here are some reasons why the number of L1 cache misses may not be equal to the number of L2 cache accesses.
You don't say what systems you are working on - I know my answer is applicable to Intel x86s such as Nehalem and Sandybridge, whose EMON performance event monitoring allows you to count things such as L1 and L2 cache misses, etc. It will probably also apply to any modern microprocessor with hardware performance counters for cache misses, such as those on ARM and Power.
Most modern microprocessors do not stop at the first cache miss, but keep going trying to do extra work. This is overall often called speculative execution. Furthermore, the processor may be in-order or out-of-order, but although the latter may given you even greater differences between number of L1 misses and number of L2 accesses, it's not necessary - you can get this behavior even on in-order processors.
Short answer: many of these speculative memory accesses will be to the same memory location. They will be squashed and combined.
The performance event "L1 cache misses" is probably[*] counting the number of (speculative) instructions that missed the L1 cache. Which then allocate a hardware data structure, called at Intel a fill buffer, at some other places a miss status handling register. Subsequent cache misses that are to the same cache line will miss the L1 cache but hit the fill buffer, and will get squashed. Only one of them, typically the first will get sent to the L2, and counted as an L2 access.)
By the way, there may be a performance event for this: Squashed_Cache_Misses.
There may also be a performance event L1_Cache_Misses_Retired. But this may undercount, since speculation may pull the data into the cache, and a cache miss at retirement may never occur.
([*] By the way, when I say "probably" here I mean "On the machines that I helped design". Almost definitely. I might have to check the definition, look at the RTL, but I would be immensely surprised if not. It is almost guaranteed.)
E.g. imagine that you are accessing bytes A[0], A[1], A[2], ... A[63], A[64], ...
If the address of A[0] is equal to zero modulo 64, then A[0]..A[63] will be in the same cache line, on a machine with 64 byte cache lines. If the code that uses these is simple, it is quite possible that all of them can be issued speculatively. QED: 64 speculative memory access, 64 L1 cache misses, but only one L2 memory access.
(By the way, don't expect the numbers to be quite so clean. You might not get exactly 64 L1 accesses per L2 access.)
Some more possibilities:
If the number of L2 accesses is greater than the number of L1 cache misses (I have almost never seen it, but it is possible) you may have a memory access pattern that is confusing a hardware prefetcher. The hardware prefetcher tries to predict which cache lines you are going to need. If the prefetcher predicts badly, it may fetch cache lines that you don't actually need. Oftentimes there is a performance evernt to count Prefetches_from_L2 or Prefetches_from_Memory.
Some machines may cancel speculative accesses that have caused an L1 cache miss, before they are sent to the L2. However, I don't know of Intel doing this.
The write policy of a data cache determines whether a store hit writes its data only on that cache (write-back or copy-back) or also at the following level of the cache hierarchy (write-through).
Hence, a store that hits at a write-through L1-D cache, also writes its data at the L2 cache.
This could be another source of L2 accesses that do not come from L1 cache misses.

Why is the size of L1 cache smaller than that of the L2 cache in most of the processors?

Why is the size of L1 cache smaller than that of the L2 cache in most of the processors ?
L1 is very tightly coupled to the CPU core, and is accessed on every memory access (very frequent). Thus, it needs to return the data really fast (usually within on clock cycle). Latency and throughput (bandwidth) are both performance-critical for L1 data cache. (e.g. four cycle latency, and supporting two reads and one write by the CPU core every clock cycle). It needs lots of read/write ports to support this high access bandwidth. Building a large cache with these properties is impossible. Thus, designers keep it small, e.g. 32KB in most processors today.
L2 is accessed only on L1 misses, so accesses are less frequent (usually 1/20th of the L1). Thus, L2 can have higher latency (e.g. from 10 to 20 cycles) and have fewer ports. This allows designers to make it bigger.
L1 and L2 play very different roles. If L1 is made bigger, it will increase L1 access latency which will drastically reduce performance because it will make all dependent loads slower and harder for out-of-order execution to hide. L1 size is barely debatable.
If we removed L2, L1 misses will have to go to the next level, say memory. This means that a lot of access will be going to memory which would imply we need more memory bandwidth, which is already a bottleneck. Thus, keeping the L2 around is favorable.
Experts often refer to L1 as a latency filter (as it makes the common case of L1 hits faster) and L2 as a bandwidth filter as it reduces memory bandwidth usage.
Note: I have assumed a 2-level cache hierarchy in my argument to make it simpler. In many of today's multicore chips, there's an L3 cache shared between all the cores, while each core has its own private L1 and maybe L2. In these chips, the shared last-level cache (L3) plays the role of memory bandwidth filter. L2 plays the role of on-chip bandwidth filter, i.e. it reduces access to the on-chip interconnect and the L3. This allows designers to use a lower-bandwidth interconnect like a ring, and a slow single-port L3, which allows them to make L3 bigger.
Perhaps worth mentioning that the number of ports is a very important design point because it affects how much chip area the cache consumes. Ports add wires to the cache which consumes a lot of chip area and power.
There are different reasons for that.
L2 exists in the system to speedup the case where there is a L1 cache miss. If the size of L1 was the same or bigger than the size of L2, then L2 could not accomodate for more cache lines than L1, and would not be able to deal with L1 cache misses. From the design/cost perspective, L1 cache is bound to the processor and faster than L2. The whole idea of caches is that you speed up access to the slower hardware by adding intermediate hardware that is more performing (and expensive) than the slowest hardware and yet cheaper than the faster hardware you have. Even if you decided to double the L1 cache, you would also increment L2, to speedup L1-cache misses.
So why is there L2 cache at all? Well, L1 cache is usually more performant and expensive to build, and it is bound to a single core. This means that increasing the L1 size by a fixed quantity will have that cost multiplied by 4 in a dual core processor, or by 8 in a quad core. L2 is usually shared by different cores --depending on the architecture it can be shared across a couple or all cores in the processor, so the cost of increasing L2 would be smaller even if the price of L1 and L2 were the same --which it is not.
#Aater's answer explains some of the basics. I'll add some more details + an examples of the real cache organization on Intel Haswell and AMD Piledriver, with latencies and other properties, not just size.
For some details on IvyBridge, see my answer on "How can cache be that fast?", with some discussion of the overall load-use latency including address-calculation time, and widths of the data busses between different levels of cache.
L1 needs to be very fast (latency and throughput), even if that means a limited hit-rate. L1d also needs to support single-byte stores on almost all architectures, and (in some designs) unaligned accesses. This makes it hard to use ECC (error correction codes) to protect the data, and in fact some L1d designs (Intel) just use parity, with better ECC only in outer levels of cache (L2/L3) where the ECC can be done on larger chunks for lower overhead.
It's impossible to design a single level of cache that could provide the low average request latency (averaged over all hits and misses) of a modern multi-level cache. Since modern systems have multiple very hungry cores all sharing a connection to the same relatively-high latency DRAM, this is essential.
Every core needs its own private L1 for speed, but at least the last level of cache is typically shared, so a multi-threaded program that reads the same data from multiple threads doesn't have to go to DRAM for it on each core. (And to act as a backstop for data written by one core and read by another). This requires at least two levels of cache for a sane multi-core system, and is part of the motivation for more than 2 levels in current designs. Modern multi-core x86 CPUs have a fast 2-level cache in each core, and a larger slower cache shared by all cores.
L1 hit-rate is still very important, so L1 caches are not as small / simple / fast as they could be, because that would reduce hit rates. Achieving the same overall performance would thus require higher levels of cache to be faster. If higher levels handle more traffic, their latency is a bigger component of the average latency, and they bottleneck on their throughput more often (or need higher throughput).
High throughput often means being able to handle multiple reads and writes every cycle, i.e. multiple ports. This takes more area and power for the same capacity as a lower-throughput cache, so that's another reason for L1 to stay small.
L1 also uses speed tricks that wouldn't work if it was larger. i.e. most designs use Virtually-Indexed, Physically Tagged (VIPT) L1, but with all the index bits coming from below the page offset so they behave like PIPT (because the low bits of a virtual address are the same as in the physical address). This avoids synonyms / homonyms (false hits or the same data being in the cache twice, and see Paul Clayton's detailed answer on the linked question), but still lets part of the hit/miss check happen in parallel with the TLB lookup. A VIVT cache doesn't have to wait for the TLB, but it has to be invalidated on every change to the page tables.
On x86 (which uses 4kiB virtual memory pages), 32kiB 8-way associative L1 caches are common in modern designs. The 8 tags can be fetched based on the low 12 bits of the virtual address, because those bits are the same in virtual and physical addresses (they're below the page offset for 4kiB pages). This speed-hack for L1 caches only works if they're small enough and associative enough that the index doesn't depend on the TLB result. 32kiB / 64B lines / 8-way associativity = 64 (2^6) sets. So the lowest 6 bits of an address select bytes within a line, and the next 6 bits index a set of 8 tags. This set of 8 tags is fetched in parallel with the TLB lookup, so the tags can be checked in parallel against the physical-page selection bits of the TLB result to determine which (if any) of the 8 ways of the cache hold the data. (Minimum associativity for a PIPT L1 cache to also be VIPT, accessing a set without translating the index to physical)
Making a larger L1 cache would mean it had to either wait for the TLB result before it could even start fetching tags and loading them into the parallel comparators, or it would have to increase in associativity to keep log2(sets) + log2(line_size) <= 12. (More associativity means more ways per set => fewer total sets = fewer index bits). So e.g. a 64kiB cache would need to be 16-way associative: still 64 sets, but each set has twice as many ways. This makes increasing L1 size beyond the current size prohibitively expensive in terms of power, and probably even latency.
Spending more of your power budget on L1D cache logic would leave less power available for out-of-order execution, decoding, and of course L2 cache and so on. Getting the whole core to run at 4GHz and sustain ~4 instructions per clock (on high-ILP code) without melting requires a balanced design. See this article: Modern Microprocessors: A 90-Minute Guide!.
The larger a cache is, the more you lose by flushing it, so a large VIVT L1 cache would be worse than the current VIPT-that-works-like-PIPT. And a larger but higher-latency L1D would probably also be worse.
According to #PaulClayton, L1 caches often fetch all the data in a set in parallel with the tags, so it's there ready to be selected once the right tag is detected. The power cost of doing this scales with associativity, so a large highly-associative L1 would be really bad for power-use as well as die-area (and latency). (Compared to L2 and L3, it wouldn't be a lot of area, but physical proximity is important for latency. Speed-of-light propagation delays matter when clock cycles are 1/4 of a nanosecond.)
Slower caches (like L3) can run at a lower voltage / clock speed to make less heat. They can even use different arrangements of transistors for each storage cell, to make memory that's more optimized for power than for high speed.
There are a lot of power-use related reasons for multi-level caches. Power / heat is one of the most important constraints in modern CPU design, because cooling a tiny chip is hard. Everything is a tradeoff between speed and power (and/or die area). Also, many CPUs are powered by batteries or are in data-centres that need extra cooling.
L1 is almost always split into separate instruction and data caches. Instead of an extra read port in a unified L1 to support code-fetch, we can have a separate L1I cache tied to a separate I-TLB. (Modern CPUs often have an L2-TLB, which is a second level of cache for translations that's shared by the L1 I-TLB and D-TLB, NOT a TLB used by the regular L2 cache). This gives us 64kiB total of L1 cache, statically partitioned into code and data caches, for much cheaper (and probably lower latency) than a monster 64k L1 unified cache with the same total throughput. Since there is usually very little overlap between code and data, this is a big win.
L1I can be placed physically close to the code-fetch logic, while L1D can be physically close to the load/store units. Speed-of-light transmission-line delays are a big deal when a clock cycle lasts only 1/3rd of a nanosecond. Routing the wiring is also a big deal: e.g. Intel Broadwell has 13 layers of copper above the silicon.
Split L1 helps a lot with speed, but unified L2 is the best choice.
Some workloads have very small code but touch lots of data. It makes sense for higher-level caches to be unified to adapt to different workloads, instead of statically partitioning into code vs. data. (e.g. almost all of L2 will be caching data, not code, while running a big matrix multiply, vs. having a lot of code hot while running a bloated C++ program, or even an efficient implementation of a complicated algorithm (e.g. running gcc)). Code can be copied around as data, not always just loaded from disk into memory with DMA.
Caches also need logic to track outstanding misses (since out-of-order execution means that new requests can keep being generated before the first miss is resolved). Having many misses outstanding means you overlap the latency of the misses, achieving higher throughput. Duplicating the logic and/or statically partitioning between code and data in L2 would not be good.
Larger lower-traffic caches are also a good place to put pre-fetching logic. Hardware pre-fetching enables good performance for things like looping over an array without every piece of code needing software-prefetch instructions. (SW prefetch was important for a while, but HW prefetchers are smarter than they used to be, so that advice in Ulrich Drepper's otherwise excellent What Every Programmer Should Know About Memory is out-of-date for many use cases.)
Low-traffic higher level caches can afford the latency to do clever things like use an adaptive replacement policy instead of the usual LRU. Intel IvyBridge and later CPUs do this, to resist access patterns that get no cache hits for a working set just slightly too large to fit in cache. (e.g. looping over some data in the same direction twice means it probably gets evicted just before it would be reused.)
A real example: Intel Haswell. Sources: David Kanter's microarchitecture analysis and Agner Fog's testing results (microarch pdf). See also Intel's optimization manuals (links in the x86 tag wiki).
Also, I wrote up a separate answer on: Which cache mapping technique is used in intel core i7 processor?
Modern Intel designs use a large inclusive L3 cache shared by all cores as a backstop for cache-coherence traffic. It's physically distributed between the cores, with 2048 sets * 16-way (2MiB) per core (with an adaptive replacement policy in IvyBridge and later).
The lower levels of cache are per-core.
L1: per-core 32kiB each instruction and data (split), 8-way associative. Latency = 4 cycles. At least 2 read ports + 1 write port. (Maybe even more ports to handle traffic between L1 and L2, or maybe receiving a cache line from L2 conflicts with retiring a store.) Can track 10 outstanding cache misses (10 fill buffers).
L2: unified per-core 256kiB, 8-way associative. Latency = 11 or 12 cycles. Read bandwidth: 64 bytes / cycle. The main prefetching logic prefetches into L2. Can track 16 outstanding misses. Can supply 64B per cycle to the L1I or L1D. Actual port counts unknown.
L3: unified, shared (by all cores) 8MiB (for a quad-core i7). Inclusive (of all the L2 and L1 per-core caches). 12 or 16 way associative. Latency = 34 cycles. Acts as a backstop for cache-coherency, so modified shared data doesn't have to go out to main memory and back.
Another real example: AMD Piledriver: (e.g. Opteron and desktop FX CPUs.) Cache-line size is still 64B, like Intel and AMD have used for several years now. Text mostly copied from Agner Fog's microarch pdf, with additional info from some slides I found, and more details on the write-through L1 + 4k write-combining cache on Agner's blog, with a comment that only L1 is WT, not L2.
L1I: 64 kB, 2-way, shared between a pair of cores (AMD's version of SMD has more static partitioning than Hyperthreading, and they call each one a core. Each pair shares a vector / FPU unit, and other pipeline resources.)
L1D: 16 kB, 4-way, per core. Latency = 3-4 c. (Notice that all 12 bits below the page offset are still used for index, so the usual VIPT trick works.) (throughput: two operations per clock, up to one of them being a store). Policy = Write-Through, with a 4k write-combining cache.
L2: 2 MB, 16-way, shared between two cores. Latency = 20 clocks. Read throughput 1 per 4 clock. Write throughput 1 per 12 clock.
L3: 0 - 8 MB, 64-way, shared between all cores. Latency = 87 clock. Read throughput 1 per 15 clock. Write throughput 1 per 21 clock
Agner Fog reports that with both cores of a pair active, L1 throughput is lower than when the other half of a pair is idle. It's not known what's going on, since the L1 caches are supposed to be separate for each core.
The other answers here give specific and technical reasons why L1 and L2 are sized as they are, and while many of them are motivating considerations for particular architectures, they aren't really necessary: the underlying architectural pressure leading to increasing (private) cache sizes as you move away from the core is fairly universal and is the same as the reasoning for multiple caches in the first place.
The three basic facts are:
The memory accesses for most applications exhibit a high degree of temporal locality, with a non-uniform distribution.
Across a large variety of process and designs, cache size and cache speed (latency and throughput) can be traded off against each other1.
Each distinct level of cache involves incremental design and performance cost.
So at a basic level, you might be able to say double the size of the cache, but incur a latency penalty of 1.4 compared to the smaller cache.
So it becomes an optimization problem: how many caches should you have and how large should they be? If memory access was totally uniform within the working set size, you'd probably end up with a single fairly large cache, or no cache at all. However, access is strongly non-uniform, so a small-and-fast cache can capture a large number of accesses, disproportionate to it's size.
If fact 2 didn't exist, you'd just create a very big, very fast L1 cache within the other constraints of your chip and not need any other cache levels.
If fact 3 didn't exist, you'd end up with a huge number of fine-grained "caches", faster and small at the center, and slower and larger outside, or perhaps a single cache with variable access times: faster for the parts closest to the core. In practice, rule 3 means that each level of cache has an additional cost, so you usually end up with a few quantized levels of cache2.
Other Constraints
This gives a basic framework to understand cache count and cache sizing decisions, but there are secondary factors at work as well. For example, Intel x86 has 4K page sizes and their L1 caches use a VIPT architecture. VIPT means that the size of the cache divided by the number of ways cannot be larger3 than 4 KiB. So an 8-way L1 cache as used on the half dozen Intel designs can be at most 4 KiB * 8 = 32 KiB. It is probably no coincidence that that's exactly the size of the L1 cache on those designs! If it weren't for this constraint, it is entirely possible you'd have seen lower-associativity and/or larger L1 caches (e.g., 64 KiB, 4-way).
1 Of course, there are other factors involved in the tradeoff as well, such as area and power, but holding those factors constant the size-speed tradeoff applies, and even if not held constant the basic behavior is the same.
2 In addition to this pressure, there is a scheduling benefit to known-latency caches, like most L1 designs: and out-of-order scheduler can optimistically submit operations that depend on a memory load on the cycle that the L1 cache would return, reading the result off the bypass network. This reduces contention and perhaps shaves a cycle of latency off the critical path. This puts some pressure on the innermost cache level to have uniform/predictable latency and probably results in fewer cache levels.
3 In principle, you can use VIPT caches without this restriction, but only by requiring OS support (e.g., page coloring) or with other constraints. The x86 arch hasn't done that and probably can't start now.
For those interested in this type of questions, my university recommends Computer Architecture: A Quantitative Approach and Computer Organization and Design: The Hardware/Software Interface. Of course, if you don't have time for this, a quick overview is available on Wikipedia.
I think the main reason for this is, that L1-Cache is faster and so it's more expensive.
https://en.wikichip.org/wiki/amd/microarchitectures/zen#Die
Compare the size of the L1, L2, and L3 caches physical size for an AMD Zen core, for example. The density increases dramatically with the cache level.
logically, the question answers itself.
If L1 were bigger than L2 (combined), then there would be no need of L2 Cache.
Why would you store your stuff on tape-drive if you can store all of it on HDD ?

Resources