There are a few things wrong with this benchmark. In Lush, he's measuring the time it takes to sum 3 million pre-computed random numbers. In Clojure, he's measu

There are a few things wrong with this benchmark. In Lush, he's measuring the time it takes to sum 3 million pre-computed random numbers. In Clojure, he's measuring the time it takes to generate 3 million random numbers, and then sum them. It also looks like he's using two different data structures. In Clojure, its a lazily generated linked list of objects; in Lush, it appears as if he's using a vector pre-initiated to the right size. He also doesn't mention which versions of Clojure and Lush he's using, or how much memory Lush uses. He complains that a JVM with a 130M is too small, but the 30 million long array in Lush would have been almost twice that size if it was populated with 64-bit doubles. For reference, it takes me 571ms to run the 300 thousand number example that took him 861ms. If I factor out the random-number generation from the Clojure benchmark, it cuts the time down to 120ms. If then I use a pre-generated fast array and the specialized areduce function (to match his specialized idx-sum function), I get 16ms for 300k, 22ms for 3m and 230ms for 30m. So performance is on par with Lush, so long as you compare like for like.