size_t WeightedCh3HashFunc::operator()(folly::StringPiece key) const {
auto n = weights_.size();
checkLogic(n && n <= furc_maximum_pool_size(), "Invalid pool size: {}", n);
size_t salt = 0;
size_t index = 0;
std::string saltedKey;
auto originalKey = key;
for (size_t i = 0; i < kNumTries; ++i) {
index = furc_hash(key.data(), key.size(), n);
/* Use 32-bit hash, but store in 64-bit ints so that
we don't have to deal with overflows */
uint64_t p = folly::hash::SpookyHashV2::Hash32(key.data(), key.size(),
kHashSeed);
assert(0 <= weights_[index] && weights_[index] <= 1.0);
uint64_t w = weights_[index] * std::numeric_limits<uint32_t>::max();
/* Rehash only if p is out of range */
if (LIKELY(p < w)) {
return index;
}
/* Change the key to rehash */
auto s = salt++;
saltedKey = originalKey.str();
do {
saltedKey.push_back(char(s % 10) + '0');
s /= 10;
} while (s > 0);
key = saltedKey;
}
return index;
}
uint32_t furc_hash_internal(const char* const key, const size_t len, const uint32_t m) {
uint32_t tries;
uint32_t d;
uint32_t num;
uint32_t i;
uint32_t a;
uint64_t hash[FURC_CACHE_SIZE];
int32_t old_ord;
assert(m <= furc_maximum_pool_size());
if (m <= 1) {
return 0;
}
furc_get_bit(NULL, 0, 0, hash, &old_ord);
for (d = 0; m > (1ul << d); d++)
;
a = d;
for (tries = 0; tries < MAX_TRIES; tries++) {
while (!furc_get_bit(key, len, a, hash, &old_ord)) {
if (--d == 0) {
return 0;
}
a = d;
}
a += FURC_SHIFT;
num = 1;
for (i = 0; i < d-1; i++) {
num = (num << 1) | furc_get_bit(key, len, a, hash, &old_ord);
a += FURC_SHIFT;
}
if (num < m) {
return num;
}
}
// Give up; return 0, which is a legal value in all cases.
return 0;
}
/**
* This verifies that
* 1) the load is evenly balanced across servers.
* 2) the act of adding a server to a pool will never result in a server
* handling keyspace that it previously handled but no longer does.
* If this occurs, then stale data may be returned.
*/
TEST(ch3, verify_correctness) {
uint32_t i, j;
uint32_t maximum_pool_size = furc_maximum_pool_size();
char key[MAX_KEY_LENGTH + 1];
std::vector<uint64_t> pools[NUM_POOLS];
uint32_t sizes[NUM_POOLS];
size_t num_pools;
auto weights = std::make_unique<std::array<double, 1U << 23U>>();
weights->fill(1.0);
srand(time(nullptr));
for (num_pools = 0; /* see end of loop */; ++num_pools) {
if (num_pools == 0) {
sizes[num_pools] = 1;
} else if (num_pools == NUM_POOLS - 1) {
sizes[num_pools] = maximum_pool_size;
} else if (num_pools % 2 == 1) { // grow pool size geometrically
sizes[num_pools] = sizes[num_pools - 1] * drand_in_range(1.5, 2.5);
} else { // grow pool size arithmetically
sizes[num_pools] = sizes[num_pools - 1] + rand_in_range(1, 11);
}
/* Make sure we don't exceed the maximum pool size. */
if (sizes[num_pools] > maximum_pool_size) {
sizes[num_pools] = maximum_pool_size;
}
pools[num_pools] = std::vector<uint64_t>(sizes[num_pools]);
if (sizes[num_pools] == maximum_pool_size)
break;
}
for (i = 0; i < NUM_SAMPLES; ++i) {
size_t previous_num = -1;
int len;
make_random_key(key, MAX_KEY_LENGTH);
len = strlen(key);
// hash the same key in each pool, in increasing pool size order
for (j = 0; j < num_pools; ++j) {
size_t num = furc_hash(key, len, sizes[j]);
EXPECT_LT(num, sizes[j]);
// Verify that the weighted furc yields identical result with weights at 1
assert(sizes[j] <= weights->size());
folly::Range<const double*> weightRange(
weights->cbegin(), weights->cbegin() + sizes[j]);
size_t weighted = facebook::mcrouter::weightedFurcHash(
folly::StringPiece(key, len), weightRange);
EXPECT_EQ(num, weighted);
++pools[j][num];
// make sure that this key either hashes the same server,
// or hashes to a new server
if (previous_num != num && j > 0) {
EXPECT_GE(num, sizes[j - 1]);
}
previous_num = num;
}
}
for (i = 0; i < num_pools; ++i) {
/* Verify that load is evenly distributed. This isn't easy to do
generally without significantly increasing the runtime by choosing
a huge NUM_SAMPLES, so just check pools up to 1000 in size. */
uint32_t pool_size = sizes[i];
if (pool_size > 1000)
break;
double expected_mean = ((double)NUM_SAMPLES) / pool_size;
double max_diff = 0;
double sum = 0;
for (j = 0; j < pool_size; j++) {
double diff = std::abs(pools[i][j] - expected_mean);
if (diff > max_diff)
max_diff = diff;
sum += pools[i][j];
}
double mean = sum / pool_size;
// expect the sample mean to be within 5% of expected mean
EXPECT_NEAR(mean, expected_mean, expected_mean * 0.05);
// expect the maximum deviation from mean to be within 15%
EXPECT_NEAR(max_diff, 0, mean * 0.15);
sum = 0;
for (j = 0; j < pool_size; j++) {
double diff = pools[i][j] - mean;
sum += diff * diff;
}
double stddev = sqrt(sum / pool_size);
// expect the standard deviation to be < 5%
EXPECT_NEAR(stddev, 0, mean * 0.05);
}
}