diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index e572278..14f9e0a 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -11,3 +11,8 @@ add_subdirectory(
     ${CMAKE_CURRENT_LIST_DIR}/simple-mallocMC
     ${CMAKE_BINARY_DIR}/examples/simple-mallocMC
 )
+
+add_subdirectory(
+    ${CMAKE_CURRENT_LIST_DIR}/multi-setup
+    ${CMAKE_BINARY_DIR}/examples/multi-setup
+)
diff --git a/examples/multi-setup/CMakeLists.txt b/examples/multi-setup/CMakeLists.txt
new file mode 100644
index 0000000..0f773df
--- /dev/null
+++ b/examples/multi-setup/CMakeLists.txt
@@ -0,0 +1,50 @@
+cmake_minimum_required(VERSION 3.14...3.22)
+
+if(POLICY CMP0167)
+  cmake_policy(SET CMP0167 NEW)
+endif()
+project(KitGenBenchExampleMultiSetup LANGUAGES CXX)
+
+# --- Import tools ----
+
+include(../../cmake/tools.cmake)
+
+# ---- Dependencies ----
+
+include(../../cmake/CPM.cmake)
+
+cpmaddpackage(
+  NAME nlohmann_json
+  GITHUB_REPOSITORY nlohmann/json
+  VERSION 3.11.3 NO_TESTS
+)
+
+cpmaddpackage(
+  NAME alpaka
+  GITHUB_REPOSITORY alpaka-group/alpaka
+  GIT_TAG 1.2.0
+)
+
+cpmaddpackage(NAME KitGenBench SOURCE_DIR ${CMAKE_CURRENT_LIST_DIR}/../..)
+
+# ---- Create standalone executable ----
+
+file(GLOB sources CONFIGURE_DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/source/*.cpp)
+
+alpaka_add_executable(${PROJECT_NAME} ${sources})
+
+set_target_properties(
+    ${PROJECT_NAME}
+    PROPERTIES
+        CXX_STANDARD 20
+        OUTPUT_NAME ${PROJECT_NAME}
+        CXX_STANDARD_REQUIRED ON
+        CXX_EXTENSIONS OFF
+)
+
+target_link_libraries(
+    ${PROJECT_NAME}
+    KitGenBench::KitGenBench
+    nlohmann_json::nlohmann_json
+    alpaka::alpaka
+)
diff --git a/examples/multi-setup/source/main.cpp b/examples/multi-setup/source/main.cpp
new file mode 100644
index 0000000..ae3e22d
--- /dev/null
+++ b/examples/multi-setup/source/main.cpp
@@ -0,0 +1,320 @@
+#include <kitgenbench/DeviceClock.h>
+#include <kitgenbench/kitgenbench.h>
+#include <kitgenbench/setup.h>
+#include <kitgenbench/version.h>
+#include <sys/types.h>
+
+#include <alpaka/workdiv/WorkDivMembers.hpp>
+#include <cstdint>
+#include <limits>
+#include <sstream>
+#include <tuple>
+#include <utility>
+#include <variant>
+#ifdef ALPAKA_ACC_GPU_CUDA_ENABLED
+#  include <cuda_runtime.h>
+#endif  //  alpaka_ACC_GPU_CUDA_ENABLE
+
+#include <alpaka/acc/AccCpuSerial.hpp>
+#include <alpaka/acc/Tag.hpp>
+#include <alpaka/atomic/Traits.hpp>
+#include <alpaka/core/Common.hpp>
+#include <alpaka/mem/buf/Traits.hpp>
+#include <alpaka/mem/view/Traits.hpp>
+#include <cstdlib>
+#include <type_traits>
+#include <vector>
+
+#include "nlohmann/json_fwd.hpp"
+
+using nlohmann::json;
+using namespace kitgenbench;
+
+using Dim = alpaka::DimInt<1>;
+using Idx = std::uint32_t;
+using AccTag = std::remove_cvref_t<decltype(std::get<0>(alpaka::EnabledAccTags{}))>;
+using Acc = alpaka::TagToAcc<AccTag, Dim, Idx>;
+
+namespace kitgenbench::Actions {
+  [[maybe_unused]] static constexpr int MALLOC = 1;
+  [[maybe_unused]] static constexpr int FREE = 2;
+}  // namespace kitgenbench::Actions
+
+auto makeExecutionDetails() {
+  auto const platformAcc = alpaka::Platform<Acc>{};
+  auto const dev = alpaka::getDevByIdx(platformAcc, 0);
+#ifdef ALPAKA_ACC_GPU_CUDA_ENABLED
+  cudaDeviceSetLimit(cudaLimitMallocHeapSize, 1024U * 1024U * 1024U);
+#endif
+  uint32_t const numThreadsPerBlock = 256U;
+  uint32_t const numThreads = 4U * numThreadsPerBlock;
+  auto workdiv = [numThreads, numThreadsPerBlock]() -> alpaka::WorkDivMembers<Dim, Idx> {
+    if constexpr (std::is_same_v<alpaka::AccToTag<Acc>, alpaka::TagCpuSerial>) {
+      return {{1U}, {1U}, {numThreads}};
+    } else {
+      return alpaka::WorkDivMembers<Dim, Idx>{
+          {numThreads / numThreadsPerBlock}, {numThreadsPerBlock}, {1U}};
+    }
+  }();
+  return kitgenbench::ExecutionDetails<Acc, decltype(dev)>{workdiv, dev};
+}
+
+// Reasons for the check to yield the result it yielded.
+// `completed` means that the check completed. The result can still be true/false depending on
+// whether the obtained value was actually correct. `notApplicable` means that the checks were
+// skipped. `nullpointer` means that a nullpointer was given, so the checks couldn't run at all.
+enum class Reason { completed, notApplicable, nullpointer };
+using Payload = std::variant<std::span<std::byte>, std::pair<bool, Reason>>;
+
+template <typename TAccTag> struct SimpleSumLogger {
+  using Clock = DeviceClock<TAccTag>;
+
+  DeviceClock<TAccTag>::DurationType mallocDuration;
+  std::uint32_t mallocCounter{0U};
+
+  DeviceClock<TAccTag>::DurationType freeDuration;
+  std::uint32_t freeCounter{0U};
+
+  std::uint32_t nullpointersObtained{0U};
+  std::uint32_t failedChecksCounter{0U};
+  std::uint32_t invalidCheckResults{0U};
+
+  template <typename TAcc> ALPAKA_FN_INLINE ALPAKA_FN_ACC auto call(TAcc const& acc, auto func) {
+    static_assert(
+        std::is_same_v<alpaka::TagToAcc<TAccTag, alpaka::Dim<Acc>, alpaka::Idx<Acc>>, TAcc>);
+    auto start = Clock::clock();
+    auto result = func(acc);
+    auto end = Clock::clock();
+
+    if (std::get<0>(result) == Actions::MALLOC) {
+      mallocDuration += Clock::duration(start, end);
+      mallocCounter++;
+    }
+
+    if (std::get<0>(result) == Actions::FREE) {
+      freeDuration += Clock::duration(start, end);
+      freeCounter++;
+    }
+
+    if (std::get<0>(result) == Actions::CHECK) {
+      if (std::holds_alternative<std::pair<bool, Reason>>(std::get<1>(result))) {
+        auto [passed, reason] = std::get<std::pair<bool, Reason>>(std::get<1>(result));
+        if (not passed) {
+          if (reason == Reason::nullpointer) {
+            nullpointersObtained++;
+          }
+          if (reason == Reason::completed) {
+            failedChecksCounter++;
+          }
+        }
+      } else {
+        invalidCheckResults++;
+      }
+    }
+
+    return result;
+  }
+
+  ALPAKA_FN_ACC void accumulate(const auto& acc, const SimpleSumLogger& other) {
+    alpaka::atomicAdd(acc, &mallocDuration, other.mallocDuration);
+    alpaka::atomicAdd(acc, &mallocCounter, other.mallocCounter);
+    alpaka::atomicAdd(acc, &freeDuration, other.freeDuration);
+    alpaka::atomicAdd(acc, &freeCounter, other.freeCounter);
+    alpaka::atomicAdd(acc, &nullpointersObtained, other.nullpointersObtained);
+    alpaka::atomicAdd(acc, &failedChecksCounter, other.failedChecksCounter);
+    alpaka::atomicAdd(acc, &invalidCheckResults, other.invalidCheckResults);
+  }
+
+  nlohmann::json generateReport() {
+#ifdef ALPAKA_ACC_GPU_CUDA_ENABLED
+    cudaDeviceProp prop;
+    cudaGetDeviceProperties(&prop, 0);
+    auto clockRate = prop.clockRate;
+#else
+    auto clockRate = 1;
+#endif  // ALPAKA_ACC_GPU_CUDA_ENABLED
+    return {
+#ifdef ALPAKA_ACC_GPU_CUDA_ENABLED
+        {"clock rate [1/ms]", clockRate},
+#endif
+        {"allocation total time [ms]", mallocDuration / clockRate},
+        {"allocation average time [ms]",
+         mallocDuration / clockRate / (mallocCounter > 0 ? mallocCounter : 1U)},
+        {"allocation count", mallocCounter},
+        {"deallocation total time [ms]", freeDuration / clockRate},
+        {"deallocation average time [ms]",
+         freeDuration / clockRate / (freeCounter > 0 ? freeCounter : 1U)},
+        {"deallocation count ", freeCounter},
+        {"failed checks count", failedChecksCounter},
+        {"nullpointers count", nullpointersObtained},
+        {"invalid check results count", invalidCheckResults},
+    };
+  }
+};
+
+template <template <typename, size_t> typename T, typename TType, size_t TExtent> struct IsSpan {
+  static constexpr bool value = std::is_same_v<T<TType, TExtent>, std::span<TType, TExtent>>;
+};
+
+template <template <typename, size_t> typename T, typename TType, size_t TExtent>
+constexpr auto isSpan(T<TType, TExtent>) {
+  return IsSpan<T, TType, TExtent>{};
+}
+
+template <typename TNew, typename TOld> constexpr auto convertDataType(std::span<TOld>& range) {
+  return std::span<TNew>(reinterpret_cast<TNew*>(range.data()),
+                         range.size() * sizeof(TOld) / sizeof(TNew));
+}
+
+struct IotaReductionChecker {
+  uint32_t currentValue{};
+
+  ALPAKA_FN_ACC auto check([[maybe_unused]] const auto& acc, const auto& result) {
+    if (std::get<0>(result) != Actions::MALLOC) {
+      return std::make_tuple(Actions::CHECK, Payload(std::make_pair(true, Reason::notApplicable)));
+    }
+    auto range = std::get<0>(std::get<1>(result));
+    if (range.data() == nullptr) {
+      return std::make_tuple(Actions::CHECK, Payload(std::make_pair(false, Reason::nullpointer)));
+    }
+    auto uintRange = convertDataType<uint32_t>(range);
+    std::iota(std::begin(uintRange), std::end(uintRange), currentValue);
+    size_t n = uintRange.size();
+    // The exact formula is using size_t because n is size_t. Casting it down will oftentimes run
+    // into an overflow that the reduction encounters, too.
+    auto expected = static_cast<uint32_t>(n * currentValue + n * (n - 1) / 2) ^ currentValue;
+    currentValue ^= std::reduce(std::cbegin(uintRange), std::cend(uintRange));
+    return std::make_tuple(+Actions::CHECK,
+                           Payload(std::make_pair(expected == currentValue, Reason::completed)));
+  }
+
+  ALPAKA_FN_ACC auto accumulate(const auto& acc, const auto& other) {
+    alpaka::atomicXor(acc, &currentValue, other.currentValue);
+  }
+
+  nlohmann::json generateReport() { return {{"final value", currentValue}}; }
+};
+
+template <typename T> struct NoStoreProvider {
+  ALPAKA_FN_ACC T load(auto const) { return {}; }
+  ALPAKA_FN_ACC void store(auto const&, T&&, auto const) {}
+  nlohmann::json generateReport() { return {}; }
+};
+
+template <typename T> struct AccumulateResultsProvider {
+  T result{};
+  ALPAKA_FN_ACC T load(auto const) { return {}; }
+  ALPAKA_FN_ACC void store(const auto& acc, T&& instance, auto const) {
+    result.accumulate(acc, instance);
+  }
+  nlohmann::json generateReport() { return result.generateReport(); }
+};
+
+template <typename T> struct AcumulateChecksProvider {
+  T result{};
+  ALPAKA_FN_ACC T load(auto const threadIndex) { return {threadIndex}; }
+  ALPAKA_FN_ACC void store(const auto& acc, T&& instance, auto const) {
+    result.accumulate(acc, instance);
+  }
+  nlohmann::json generateReport() { return result.generateReport(); }
+};
+
+namespace setups {
+  template <uint32_t size> struct SingleSizeMallocRecipe {
+    static constexpr std::uint32_t allocationSize{size};
+    static constexpr std::uint32_t numAllocations{256U};
+    std::array<std::byte*, numAllocations> pointers{{}};
+    std::uint32_t counter{0U};
+
+    ALPAKA_FN_ACC auto next([[maybe_unused]] const auto& acc) {
+      if (counter >= numAllocations)
+        return std::make_tuple(
+            +kitgenbench::Actions::STOP,
+            Payload(std::span<std::byte>{static_cast<std::byte*>(nullptr), allocationSize}));
+      pointers[counter] = static_cast<std::byte*>(malloc(allocationSize));
+      auto result
+          = std::make_tuple(+kitgenbench::Actions::MALLOC,
+                            Payload(std::span<std::byte>(pointers[counter], allocationSize)));
+      counter++;
+      return result;
+    }
+
+    nlohmann::json generateReport() { return {}; }
+  };
+
+  template <uint32_t size, typename TAcc, typename TDev> struct InstructionDetails {
+    struct DevicePackage {
+      NoStoreProvider<SingleSizeMallocRecipe<size>> recipes{};
+      AccumulateResultsProvider<SimpleSumLogger<AccTag>> loggers{};
+      AcumulateChecksProvider<IotaReductionChecker> checkers{};
+    };
+
+    DevicePackage hostData{};
+    alpaka::Buf<TDev, DevicePackage, alpaka::Dim<TAcc>, alpaka::Idx<TAcc>> devicePackageBuffer;
+
+    InstructionDetails(TDev const& device)
+        : devicePackageBuffer(alpaka::allocBuf<DevicePackage, Idx>(device, 1U)) {};
+
+    auto sendTo([[maybe_unused]] TDev const& device, auto& queue) {
+      alpaka::memset(queue, devicePackageBuffer, 0U);
+      return reinterpret_cast<DevicePackage*>(alpaka::getPtrNative(devicePackageBuffer));
+    }
+    auto retrieveFrom([[maybe_unused]] TDev const& device, auto& queue) {
+      auto const platformHost = alpaka::PlatformCpu{};
+      auto const devHost = getDevByIdx(platformHost, 0);
+      auto view = alpaka::createView(devHost, &hostData, 1U);
+      alpaka::memcpy(queue, view, devicePackageBuffer);
+    }
+
+    nlohmann::json generateReport() {
+      return {{"recipes", hostData.recipes.generateReport()},
+              {"logs", hostData.loggers.generateReport()},
+              {"checks", hostData.checkers.generateReport()}};
+    }
+  };
+
+  template <typename TAcc, uint32_t size, typename TDev>
+  auto makeInstructionDetails(TDev const& device) {
+    return InstructionDetails<size, TAcc, TDev>(device);
+  }
+
+  using ALLOCATION_SIZES = std::integer_sequence<uint32_t, 16U, 32U, 128U, 256U, 512U, 1024U>;
+
+  template <uint32_t... nums> auto composeSetups(std::integer_sequence<uint32_t, nums...>) {
+    auto execution = makeExecutionDetails();
+    return std::make_tuple(
+        setup::composeSetup((std::stringstream{} << "Single size: " << nums).str(), execution,
+                            makeInstructionDetails<Acc, nums>(execution.device), {})...);
+  }
+}  // namespace setups
+
+/**
+ * @brief Compose a report from the provided metadata, configuration, and individual reports.
+ *
+ * This function takes a json object representing the metadata, a json object
+ * representing the configuration, and a json object representing the individual
+ * reports, and composes a report by merging them into a single json object.
+ * The resulting json object is returned.
+ *
+ * @param metadata The json object representing the metadata.
+ * @param config The json object representing the configuration.
+ * @param individualReports The json object representing the individual reports.
+ * @return json The json object representing the composed report.
+ */
+json composeReport(json const& metadata, json const& benchmarkReports) {
+  json report{};
+  report["metadata"] = metadata;
+  report["benchmarks"] = benchmarkReports;
+  return report;
+}
+
+void output(json const& report) { std::cout << report << std::endl; }
+
+auto main() -> int {
+  auto metadata = gatherMetadata();
+  auto setup = setups::composeSetups(setups::ALLOCATION_SIZES{});
+  auto benchmarkReports = std::apply([](auto&&... args) { return runBenchmarks(args...); }, setup);
+  auto report = composeReport(metadata, benchmarkReports);
+  output(report);
+  return EXIT_SUCCESS;
+}